Dynamic reconfiguration of brain coactivation states associated with active and lecture-based learning of university physics

Abstract

Academic institutions are increasingly adopting active learning methods to enhance educational outcomes. Using functional magnetic resonance imaging (fMRI), we investigated neurobiological differences between active learning and traditional lecture-based approaches in university physics education. Undergraduate students enrolled in an introductory physics course underwent an fMRI session before and after a 15-week semester. Coactivation pattern (CAP) analysis was used to examine the temporal dynamics of brain states across different cognitive contexts, including physics conceptual reasoning, physics knowledge retrieval, and rest. CAP results identified seven distinct brain states, with contributions from frontoparietal, somatomotor, and visuospatial networks. Among active learning students, physics learning was associated with increased engagement of a somatomotor network, supporting an embodied cognition framework, while lecture-based students demonstrated stronger engagement of a visuospatial network, consistent with observational learning. These findings suggest significant neural restructuring over a semester of physics learning, with different instructional approaches preferentially modulating distinct patterns of brain dynamics.

Introduction

Physics education research has led to the development of evidence-based instructional strategies aimed at effectively disseminating physics concepts to students, fostering deeper conceptual understanding, minimizing rote learning and non-divergent thinking, and enabling the construction of expert-like knowledge structures that enable students to apply physics principles to real-world problems^1,2. Among these methodologies, active learning has emerged as a pivotal strategy for developing a more integrative understanding of physics among students and has seen widespread adoption throughout various academic institutions³. Active learning strategies adopt a student-centered approach, encouraging students to engage in group activities and collaboratively teach one another relevant course material with minimal instructor intervention^4,5. In contrast, traditional, passive learning environments are characterized by an instructor-centered model, wherein the educator delivers course content to students, who are expected to process the material through listening, note-taking, and occasional participation during lectures at the instructor’s discretion^4,6.

Despite the numerous behavioral studies assessing the differences between active and passive learning methodologies in physics education^7,8,9,10,11, the longitudinal neurocognitive effects of these different instructional methods remain largely unexplored. These fundamentally different instructional strategies may selectively modulate distinct functional brain networks, which in turn could affect the underlying neural representations that are engaged when students interact with and internalize physics-related content and complex physics concepts. In contrast to passive learning, active learning emphasizes collaborative, hands-on engagement with learning materials, typically leading to enhanced conceptual understanding^12,13,14. Consequently, a theoretical framework that may provide insights into the anticipated differences in the neurobiological mechanisms of active versus passive physics learning is embodied cognition, which posits that cognitive processes are deeply rooted in the body’s interactions with the world^15,16. According to this framework, conceptual knowledge is grounded in sensory and motor experiences, and understanding abstract concepts involves simulating sensorimotor patterns associated with those concepts¹⁷. Furthermore, a recent meta-analytic study demonstrated that embodied cognitive approaches in learning environments correspond with improved comprehension, greater knowledge retention, and reduced cognitive load¹⁸.

Prior neuroimaging research using fMRI and EEG techniques has shown that processing action-related words and concepts activates sensorimotor regions of the brain, a phenomenon known as semantic somatotopy^{19,20,21,22,23}. These studies suggest that the recruitment of sensorimotor regions is important in the comprehension of abstract concepts by grounding them in concrete somatomotor representations. In the context of physics education, students who actively and physically interact with learning materials may develop stronger and more integrative neural activation patterns when engaging with physics concepts. For instance, Kontra et al.²⁰ demonstrated the relevance of embodied cognition in physics, found that university students who learned about angular momentum through interactive physical models exhibited greater activation in brain regions associated with action planning and motor experiences, such as the dorsal premotor cortex, primary motor/somatosensory cortex, superior parietal lobe, and supplementary motor area. Moreover, students who physically engaged with these concepts showed increased performance on tests of angular momentum. Furthermore, the involvement of sensorimotor regions in understanding abstract physics concepts may result from the neural reuse of regions initially dedicated to lower-level sensory and motor functions now supporting higher-order cognitive processes²⁴. This neural reuse is evident in studies demonstrating that regions in the parietal and premotor cortex are implicated in mental rotation tasks and spatial reasoning, both critical cognitive domains in physics learning, engaging brain networks such as the somatomotor, dorsal attention, and frontoparietal network (FPN)²⁵.

Evidence supporting embodied cognition in physics learning suggests that different teaching methods may create distinct patterns of brain network activations over time²⁰. Prior neuroimaging evidence has shown that learning involves dynamic brain network reconfiguration²⁶. Methods that capture time-varying properties of neuroimaging data are known to align more strongly with shifts in cognition and behavior than static functional connectivity (sFC) techniques²⁷. Thus, these methods may provide more informative insights into the neurobiological differences between active and passive instruction. SFC techniques analyze functional connectivity within and between brain regions while assuming that the observed connectivity patterns remain stationary throughout the entire data acquisition period^27,28. This approach overlooks the dynamic nature of brain connectivity, which often demonstrates considerable temporal variability within seconds during different cognitive tasks and learning processes²⁹. Capturing these temporal dynamics is particularly important as changes in functional connectivity over time may reflect underlying neuroplastic processes such as Hebbian learning, which posits that persistent activation of specific neural patterns in response to repeated engagement with tasks leads to intrinsic cortical restructuring and stable, recurrent states^30,31,32. Moreover, prior research on metastable brain states, defined as transient patterns of network activations that are observed at greater frequencies during resting state, suggests that higher metastability of certain brain states during rest is associated with increased pre-configuration towards tasks that rely on the underlying cognitive processes related to those neural patterns^33,34,35. Hence, by exploring temporal dynamics through dynamic functional connectivity (dFC) methods, transient patterns of network activations can be captured, identifying potential neural restructuring associated with different instructional methods over time and providing greater insights that may be overlooked by traditional static methods.

To capture these dynamic patterns, we leveraged the coactivation pattern (CAP) analytical technique, a dynamic method that aggregates similar spatial distributions of brain activity^36,37 to examine the temporal dynamics of brain networks across different cognitive contexts, including physics conceptual reasoning, physics knowledge (PK) retrieval, and rest. Brain states consistently identified across multiple cognitive contexts are likely to provide a robust representation of intrinsic functional connectivity that is more generalizable, reliable, and meaningful compared to states identified from single-task fMRI, especially when broad networks (e.g., default, frontoparietal, somatomotor, and attention networks) are of interest^38,39,40. We utilized three reproducible CAP metrics⁴¹: (1) temporal fraction, defined as the proportion of time occupied by a single state; (2) persistence, defined as the average duration a state persists before transitioning to another state; and (3) counts, defined as the frequency of CAP initiation (contiguous CAPs are considered a single initiation) across the entire scan.

In the present study, we investigated the neurobiological differences in dFC among students who completed a semester of an introductory undergraduate physics course in either active learning or traditional lecture-based classrooms. To this end, we assessed different cognitive contexts, using two in-scanner fMRI tasks, including the Force Concept Inventory, a physics-related conceptual reasoning task examining motion trajectories of objects at rest or in motion, and the PK task, a physics-related semantic retrieval task assessing general PK, as well as resting state. Across these different contexts, we examined how time (pre- versus post-instruction) and instructional approach (active learning versus lecture-based classrooms) were associated with different brain network activation patterns. After a semester of physics instruction, we expected students to show increased engagement of brain regions critical for physics learning, particularly the frontoparietal, somatomotor, and visual networks that support spatial comprehension and mathematical thinking^42,43. However, we anticipated that the method of instruction would preferentially modulate distinct neural patterns. Drawing on embodied cognition theory and principles of Hebbian learning, we hypothesized that active learning students would develop more integrated network activation patterns, evidenced by greater temporal fraction, persistence, and counts in CAPs characterized by widespread activation across multiple networks, including increased somatomotor engagement. We anticipated that this multimodal pattern would reflect how active learning facilitates comprehension of physics understanding through collaborative physical interaction with concepts. In contrast, we expected lecture-based students to show increased temporal fraction, persistence, and counts in CAPs dominated by visual network activation, reflecting how these students primarily acquire PK through observation⁴⁴. Finally, we predicted that instructional method-specific patterns would be most observed during resting state, suggesting that different teaching approaches are linked to intrinsic reconfiguration that generalizes beyond physics-related cognition during domain-specific tasks.

Results

Demographic differences between active learning and lecture-based classrooms

The study sample comprised 121 undergraduate students enrolled in a 15-week introductory physics course that employed either a modeling instruction approach (n = 61), emphasizing model building and collaborative group activities⁴⁵, or a traditional lecture-based instruction approach (n = 60). Behavioral and neuroimaging data were collected at two time periods: pre- and post-instruction. We assessed for potential demographic differences between students in the active learning classrooms (n = 61) and those in the lecture-based classrooms (n = 60) and found no significant differences in age, sex, ethnicity, household income, grade point average (GPA), and years enrolled at Florida International University (FIU) (i.e., freshman, sophomore, junior, or senior) (Supplementary Table 2). We further examined a subsample of participants with at least one functional run for all tasks and sessions, which included active learning (n = 46) and lecture-based students (n = 44), and also observed no significant demographic differences (Supplementary Table 2).

Spatial topography of coactivation patterns (CAPs)

We identified seven CAPs that dynamically varied at both pre- and post-instruction during the task and rest runs (Fig. 1); this number of brain states was selected using the elbow criterion (Supplementary Fig. 1). These patterns were identified through k-means clustering of concatenated resting state and task data (FCI and PK) from all participants with at least one functional run across all tasks and sessions (n = 90). The spatial dimensionality of the extracted time series data from each participant was reduced using the HCPex atlas⁴⁶, an extended Montreal Neuroimaging Institute (MNI) version of the Glasser et al.⁴⁷ parcellation, which includes 66 subcortical regions, resulting in 426 nodes across 22 cortical divisions. “High Amplitude” cosine similarities quantify the similarity between a cortical division from the HCPex atlas and the positive activations within a CAP cluster centroid. These values represent regions that are more active than average. Conversely, “Low Amplitude” cosine similarities measure the similarity between a cortical division and the negative activations within the cluster centroid, indicating regions with activity levels below the mean. While the “Low Amplitude” cosine similarity values are negative, for visualization purposes, the absolute values of the negative activations from each CAP cluster centroid were computed to restrict all cosine similarity values from 0 to 1. However, the negative cosine similarities are reported below to differentiate them from the cosine similarities of the “High Amplitude” results. Finally, we then characterized each CAP by the network correspondence between HCPex cortical divisions exhibiting a cosine similarity ≥0.20 to the positive activations in the CAP and 12 resting state networks from the Cole-Anticevic atlas⁴⁸ through spin permutation^49,50.

Fig. 1: Cosine similarity analysis of CAPS.

Surface plots for each CAP are paired with radar plots depicting the cosine similarity between each CAP and various HCPex cortical divisions. In these radar plots, “High Amplitude” (shown in red) represents the cosine similarity between the cortical division and the positive activations in the cluster centroid, while “Low Amplitude” (shown in blue) represents the cosine similarity between the cortical division and the negative activations in a CAP cluster centroid. The radial axis of each radar plot shows cosine similarity values ranging from 0 to 0.5. Note that during computation, the inverse values of the negative activations were calculated to restrict the cosine similarity for both positive and negative activations in a CAP from 0 to 1. At the bottom right, a depiction of each HCPex cortical division (excluding the subcortical areas) is shown. The labels for each CAP show the network correspondence of cosine similarities ≥0.20.

CAP-1 showed widespread activations across the Paracentral Midcingulate (${cosine\; similarity}[\cos (\theta )]=0.34$), MT+ Complex ($\cos (\theta )=0.31$), Dorsal Stream Visual ($\cos (\theta )=0.28$), Somatosensory regions ($\cos (\theta )=0.27$), Posterior Opercular ($\cos (\theta )=0.26$), Early Auditory ($\cos (\theta )=0.26$), Insula Frontal Opercular ($\cos (\theta )=0.25$), Superior Parietal ($\cos (\theta )=0.25$), Ventral Stream Visual ($\cos (\theta )=0.24$), and the Premotor ($\cos (\theta )=0.23$). These patterns demonstrated significant correspondence with the Somatomotor Network (${Dice\; Similarity\; Coefficient}[{DSC}]=0.4875$, $p=0.001$) and the Cingulo-Opercular Network (${DSC}=0.3006$, $p=0.023$). Dorsolateral Prefrontal Cortex ($\cos (\theta )=-0.41$), Posterior Cingulate ($\cos (\theta )=-0.35$), OrbPolaFrontal ($\cos (\theta )=-0.29$), Lateral Temporal ($\cos (\theta )=-0.26$), Inferior Frontal regions ($\cos (\theta )=-0.26$), and the Anterior Cingulate Medial Prefrontal Cortex ($\cos (\theta )=-0.25$).

CAP-2 demonstrated robust activations in the Dorsolateral Prefrontal Cortex ($\cos (\theta )=0.42$), Posterior Cingulate ($\cos (\theta )=0.34$), Inferior Parietal ($\cos (\theta )=0.27$, Lateral Temporal ($\cos (\theta )=0.26$), Inferior Frontal ($\cos (\theta )=0.24$), and the Orbitofrontal Cortex ($\cos (\theta )=0.24$). These activation patterns showed strong correspondence with both the Default Mode Network (${DSC}=0.4898$, $p=0.001$) and the Frontoparietal Network (${DSC}=0.4511$, $p=0.001$). Notable deactivations were observed in the Paracentral Midcingulate ($\cos (\theta )=-0.40$), Insula Frontal Opercular ($\cos (\theta )=-0.37$), Early Auditory ($\cos (\theta )=-0.37$), Posterior Opercular ($\cos (\theta )=-0.36$), Somatosensory Motor ($\cos (\theta )=-0.30$), and the Premotor ($\cos (\theta )=-0.22$).

CAP-3 exhibited significant activations in the Dorsolateral Prefrontal Cortex ($\cos (\theta )=0.32$), Superior Parietal ($\cos (\theta )=0.31$), Subcortical ($\cos (\theta )=0.27$), Insula Frontal Opercular ($\cos (\theta )=0.23$), Inferior Parietal ($\cos (\theta )=0.23$), and the Inferior Frontal ($\cos (\theta )=0.23$). These activation patterns aligned significantly with the Frontoparietal Network (${DSC}=0.3940$, $p=0.001$), Cingulo-Opercular Network (${DSC}=0.3287$, $p=0.01$3), and the Dorsal Attention Network (${DSC}=0.1629$, $p=0.007$). Substantial deactivations were observed in the Auditory Association regions ($\cos (\theta )=-0.37$), MT+ Complex ($\cos (\theta )=-0.32$), Somatosensory ($\cos (\theta )=-0.27$), Posterior Cingulate ($\cos (\theta )=-0.22$), Lateral Temporal ($\cos (\theta )=-0.22$), and the Temporoparietal-Occipital junction ($\cos (\theta )=-0.21$).

CAP-4 exhibited pronounced activations in the Anterior Cingulate Medial Prefrontal Cortex ($\cos (\theta )=0.38$), Dorsolateral Prefrontal Cortex ($\cos (\theta )=0.38$), and Posterior Cingulate ($\cos (\theta )=0.33$), alongside the OrbPolaFrontal ($\cos (\theta )=0.30$), Inferior Frontal ($\cos (\theta )=0.24$), Lateral Temporal ($\cos (\theta )=0.24$), and the Auditory Association ($\cos (\theta )=0.20$). These patterns showed significant correspondence with the Default Mode Network (${DSC}=0.5440$, $p=0.001$), Frontoparietal Network (${DSC}=0.3647$, $p=0.001$), and the Language Network (${DSC}=0.1726$, $p=0.015$). Strong deactivations were noted across visual processing areas, particularly in the MT+ Complex ($\cos (\theta )=-0.46$), Dorsal Stream Visual ($\cos (\theta )=-0.40$), Ventral Stream Visual ($\cos (\theta )=-0.40$), Superior Parietal ($\cos (\theta )=-0.33$), Early Visual ($\cos (\theta )=-0.32$), and the Inferior Parietal ($\cos (\theta )=-0.20$).

CAP-5 demonstrated robust activations in the Auditory Association ($\cos (\theta )=0.38$), MT+ Complex ($\cos (\theta )=0.28$), Somatosensory Motor ($\cos (\theta )=0.26$), Posterior Cingulate ($\cos (\theta )=0.24$), Lateral Temporal ($\cos (\theta )=0.22$), and the Anterior Cingulate Cortex ($\cos (\theta )=0.21$). These activation patterns aligned significantly with the Default Mode Network (${DSC}=0.4316$, $p=0.004$) and Somatomotor Network (${DSC}=0.2712,{p}=0.002$). Notable deactivations were observed in the Superior Parietal ($\cos (\theta )=-0.31$), Dorsolateral Prefrontal Cortex ($\cos (\theta )=-0.31$), Subcortical ($\cos (\theta )=-0.27$), Insula Frontal Opercular ($\cos (\theta )=-0.24$), Inferior Frontal ($\cos (\theta )=-0.23$), and the Inferior Parietal ($\cos (\theta )=-0.23$).

CAP-6 revealed extensive activations across visual processing regions, including MT+ Complex ($\cos (\theta )=0.45$), Ventral Stream Visual ($\cos (\theta )=0.42$), Dorsal Stream Visual ($\cos (\theta )=0.39$), Early Visual ($\cos (\theta )=0.34$), and the Superior Parietal ($\cos (\theta )=0.30$). These activation patterns showed strong correspondence with the Secondary Visual Network (${DSC}=0.7102$, $p=0.001$) and the Dorsal Attention Network (${DSC}=0.1406$, $p=0.036$). Pronounced deactivations were observed in Anterior Cingulate Medial Prefrontal Cortex ($\cos (\theta )=-0.40$), Insula Frontal Opercular ($\cos (\theta )=-0.38$), Early Auditory ($\cos (\theta )=-0.34$), Posterior Opercular ($\cos (\theta )=-0.31$), Paracentral Midcingulate ($\cos (\theta )=-0.24$), Auditory Association ($\cos (\theta )=-0.23$), and the Subcortical ($\cos (\theta )=-0.22$).

CAP-7 revealed strong activations in Insula Frontal Opercular ($\cos (\theta )=0.42$), Early Auditory ($\cos (\theta )=0.40$), Posterior Opercular ($\cos (\theta )=0.38$), Paracentral Midcingulate ($\cos (\theta )=0.36$), Somatosensory Motor ($\cos (\theta )=0.25$), and the Anterior Cingulate Medial Prefrontal Cortex ($\cos (\theta )=0.22$). These patterns demonstrated significant overlap with the Somatomotor Network (${DSC}=0.4679$, $p=0.001$) and the Cingulo-Opercular Network (${DSC}=0.3517$, $p=0.005$). Prominent deactivations were observed in the Ventral Stream Visual ($\cos (\theta )=-0.38$), MT+ Complex ($\cos (\theta )=-0.34$), Early Visual ($\cos (\theta )=-0.32$), Inferior Parietal ($\cos (\theta )=-0.32$), Dorsal Stream Visual ($\cos (\theta )=-0.3$1), and the Superior Parietal ($\cos (\theta )=-0.25$).

Comparison of the spatial patterns for the seven CAPs revealed complex relationships among them (Fig. 2). Notably, each CAP demonstrated significant anti-correlation with at least one other CAP. CAP-1 showed significant negative correlations with CAP-2 and CAP-4. CAP-2 was found to be strongly anti-correlated with CAP-7. CAP-3 had a significant negative correlation with CAP-5, while CAP-4 was anti-correlated with CAP-6 and CAP-7. Additionally, CAP-6 and CAP-7 exhibited a notable anti-correlation with each other.

Fig. 2: Visualizations of coactivation patterns.

A A heatmap depicts the mean activation of HCPex cortical divisions across seven identified CAPs. Red indicates higher activation, while blue represents lower activation or deactivation. B Surface plots illustrate the spatial distribution of activation (yellow) and deactivation (blue) for each CAP across left (L) and right (R) hemispheres. C A correlation matrix shows the Pearson correlation coefficients between all pairs of CAPs. The color scale ranges from purple (negative correlation) to yellow (positive correlation). Notably, each CAP demonstrated significant anti-correlation with at least one other CAP. CAP-1 showed significant negative correlations with CAP-2 (r = −0.849, p_uncorr < 0.001, p_BH < 0.001) and CAP-4 (r = −0.851, p_uncorr < 0.001, p_BH < 0.001). CAP-2 was found to be strongly anti-correlated with CAP-7 (r = −0.792, p_uncorr < 0.001, p_BH < 0.001). CAP-3 had a significant negative correlation with CAP-5 (r = −0.997, p_uncorr < 0.001, p_BH < 0.001), while CAP-4 was anti-correlated with CAP-6 (r = −0.751, p_uncorr < 0.001, p_BH < 0.001). Additionally, CAP-6 and CAP-7 exhibited a notable anti-correlation with each other (r = −0.786, p_uncorr < 0.001, p_BH < 0.001).

Main effect time (pre- vs. post-instruction)

A contrast analysis was performed on each linear mixed effect model to evaluate mean differences for each CAP metric from pre-instruction to post-instruction for each task (Fig. 3). These analyses were conducted by probing, for the main effect of time, several linear mixed-effects models (one model per temporal dynamic per CAP) that accounted for random effects due to individual differences while controlling for demographic variables including age, sex, ethnicity, household income, GPA, and years enrolled. For analyses involving CAP counts, the total number of TRs per participant was included as a covariate to account for differences in available data points due to motion scrubbing and task design. All p values were corrected for multiple comparisons using the Benjamini–Hochberg procedure.

Fig. 3: Main effect of time.

Results of the main effect of time, which were probed from the linear mixed modeling analyses, are displayed for: A Force Concept Inventory (FCI), B Physics Knowledge (PK), and C Resting State. Three CAP metrics are presented: Counts (left column), Temporal Fraction (middle column), and Persistence (right column). Each plot shows the distribution for each session (Pre-Instruction shown in pink and Post-Instruction in blue), accompanied by box plots displaying their means (black diamond) and edges representing the 25th and 75th quantiles of the distribution. Asterisks (*) denote statistical significance of mean differences prior to Benjamini–Hochberg correction.

In the FCI task, only CAP-6 showed increased temporal fraction ($\varDelta$ = 0.0140, ${p}_{{uncorr}}$ = 0.018, ${p}_{{BH}}$ = 0.191) from pre- to post-instruction. The PK task exhibited pre- to post-instruction changes with decreased temporal fraction ($\varDelta$ = 0.019, ${p}_{{uncorr}}\,$< 0.001, ${p}_{{BH}}\,$= 0.067) and counts ($\varDelta$ = 0.621, ${p}_{{uncorr}}\,$= 0.031, ${p}_{{BH}}\,$= 0.220) in CAP-5. The resting state showed the most widespread pre- to post-instruction changes, with significant alterations in three CAPs: CAP-1 showed decreased counts ($\varDelta$ = −1.942, ${p}_{{uncorr}}\,$< 0.001, ${p}_{{BH}}$ < 0.001), temporal fraction ($\varDelta$ = −0.019, ${p}_{{uncorr}}\,$< 0.001, ${p}_{{BH}}$ < 0.001), and persistence ($\varDelta$ = −0.202, ${p}_{{uncorr}}\,$= 0.036, ${p}_{{BH}}\,$= 0.099), CAP-4 showed decreased counts ($\varDelta$ = −1.090, ${p}_{{uncorr}}\,$= 0.011, ${p}_{{BH}}\,$= 0.039), temporal fraction ($\varDelta$ = −0.012, ${p}_{{uncorr}}\,$< 0.001, ${p}_{{BH}}$ = 0.001), and persistence ($\varDelta$ = −0.210, ${p}_{{uncorr}}\,$= 0.025, ${p}_{{BH}}\,$= 0.099), and CAP-7 showed increased counts ($\varDelta$ = 1.130, ${p}_{{uncorr}}\,$= 0.019, ${p}_{{BH}}\,$= 0.044), temporal fraction ($\varDelta$ = 0.014, ${p}_{{uncorr}}\,$< 0.001, ${p}_{{BH}}$ = 0.001), and persistence ($\varDelta$ = 0.203, ${p}_{{uncorr}}\,$= 0.047, ${p}_{{BH}}\,$= 0.099). The delta change values and associated p values for all tested models can be found in Supplementary Table 5.

Main effect of instruction (active learning vs. lecture-based classrooms)

Probing the linear mixed models for the main effect of instruction revealed several significant effects. For the main effect of instruction (Fig. 4), the active learning students exhibited lower persistence ($\varDelta =-0.267$, ${p}_{{uncorr}}=0.046$, ${p}_{{BH}}=0.322$) of CAP-2 compared to the lecture-based students during the FCI task. Additionally, the active learning students also showed reduced persistence in CAP-2 ($\varDelta =0.346$, ${p}_{{uncorr}}=0.014$, ${p}_{{BH}}=0.047$) and CAP-3 ($\varDelta =0.394$, ${p}_{{uncorr}}=0.007$, ${p}_{{BH}}=$ 0.047) during the PK task when compared to the lecture-based students. The delta change values and associated p values for all tested models can be found in Supplementary Table 6.

Fig. 4: Main effect of instruction.

Results of the main effect of instruction, which were probed from the linear mixed modeling analyses, are displayed for: A Force Concept Inventory (FCI), B Physics Knowledge (PK), and C Resting State. Three CAP metrics are presented: Counts (left column), Temporal Fraction (middle column), and Persistence (right column). Each plot shows the distribution for each classroom (active learning shown in purple and lecture-based in green), accompanied by box plots displaying their means (black diamond) and edges representing the 25th and 75th quantiles of the distribution. Asterisks (*) denote statistical significance of mean differences prior to Benjamini–Hochberg correction.

Interaction between time and instruction

Several significant interaction effects emerged (Fig. 5) from our linear mixed-effects models examining how instructional methods moderated changes in CAP metrics over time. These models included time (pre vs. post) and instruction type (active vs. lecture-based) as fixed effects, along with their interaction term, while controlling for demographic variables and random effects. The resulting beta coefficients represent how the change in CAP metrics from pre- to post-instruction differed between instructional groups, with positive values indicating greater increases or smaller decreases in temporal dynamics of a CAP within the active learning group compared to the lecture-based group.

Fig. 5: Interaction effects.

Interaction plots depicting changes in mean CAP counts, temporal fraction, and persistence from pre- to post-instruction for lecture-based (green) and active learning (purple) instructional methods across FCI, PK, and resting state data. Each plot represents a specific CAP (depicted in the right column) that showed a significant interaction effect prior to Benjamini–Hochberg correction: persistence of CAP-4 for FCI, counts of CAP-7 for PK, counts and temporal fraction of CAP-6, as well as counts of CAP-1, for resting state.

For the FCI task, CAP-4 showed a significant interaction with the active learning students demonstrating increased persistence ($\beta =0.549$, ${p}_{{uncorr}}=0.013$, ${p}_{{BH}}=0.093$) from pre- to post-instruction relative to the lecture-based students. For the PK task, the lecture-based students showed increased counts ($\beta =$-$1.120$, ${p}_{{uncorr}}=0.031$, ${p}_{{BH}}=0.215$) from pre- to post-instruction compared to the active learning students in CAP-7. For resting state, from pre- to post- instruction, the active learning students demonstrated increased persistence ($\beta =0.494$, ${p}_{{uncorr}}=0.009$, ${p}_{{BH}}=0.067$) relative to the lecture-based instruction students in CAP-1 while the lecture-based instruction group exhibited increased counts ($\beta =-1.860$, ${p}_{{uncorr}}=0.0484$, ${p}_{{BH}}=$ 0.338) and temporal fraction ($\beta =-0.016$, ${p}_{{uncorr}}=0.037$, ${p}_{{BH}}=$ 0.262) in CAP-6 when compared to the active learning students. The beta coefficients and associated p values for all tested models can be found in Supplementary Table 7.

Exploratory brain-behavioral analyses

Exploratory analyses of behavioral measures (i.e., FCI and PK task accuracy and reaction time) revealed a significant main effect of time (pre- to post-instruction) was observed. From pre-to post-instruction, both FCI ($\beta =0.071,{p}=0.018$) and PK accuracy ($\beta =0.001,p < 0.001$) increased, while PK reaction time (ms) decreased ($\beta =-138.00,p < 0.001$); however, FCI reaction time (ms) did not significantly change from pre- to post-instruction ($\beta =-213.00,p=0.692$. No significant main effect of instructional methodology, nor any interaction between time and instructional methodology, was observed (Supplementary Table 8).

Despite these behavioral improvements across time, no significant associations were found between the pre-to-post-instruction behavioral changes and the corresponding changes in relevant temporal dynamic metrics (i.e., CAPs that exhibited a main effect of time for the FCI and PK tasks). Specifically, the pre-to-post-instruction increase in the temporal fraction of CAP-6 was not significantly associated with an increase in FCI task accuracy. Similarly, for the PK task, neither the increased counts nor the increased temporal fraction of CAP-5 were significantly associated with the observed increase in PK task accuracy or the decrease in PK reaction time (Supplementary Table 9).

Discussion

In this study, we investigated changes in the temporal dynamics of seven distinct CAPs, examining how different physics instructional methodologies modulate brain state dynamics over time. Drawing from embodied cognition theory and principles of Hebbian learning, we hypothesized that active learning and traditional lecture-based instructional approaches would differentially influence how students process and internalize physics concepts. Using three CAP metrics (i.e., counts, temporal fraction, and persistence), we evaluated changes in undergraduate students’ neural dynamics before and after a semester of introductory physics during both task-based (i.e., FCI and PK) and resting state conditions. Our analysis revealed that the timing and type of instructional method and their interaction selectively modulate the temporal dynamics of distinct brain states. Moreover, our exploratory behavioral analyses showed that FCI and PK task accuracy increased from pre- to post-instruction, and PK reaction time decreased over time with no change in FCI reaction time. However, there were no behavioral differences between the modeling and lecture-based instruction groups for either the main effect of instruction or interaction (between time and instruction method) contrasts. Furthermore, significant differences in temporal dynamics observed from pre- to post-instruction for the FCI and PK tasks did not explain the increases in FCI task accuracy and PK task accuracy, nor the decrease in PK reaction time. Regardless of the lack of brain-behavioral associations from the exploratory analyses, supporting our hypotheses, these findings suggest that different instructional methodologies, specifically active and passive learning, result in divergent neural reconfigurations among students, potentially through differential engagement of sensorimotor and observational learning mechanisms.

CAPs analysis revealed that three of seven dynamic brain patterns demonstrated notable Somatomotor Network (SMN) activation. CAP-1 showed a multimodal pattern with increased activation in brain regions associated with the Dorsal Attention Network (DAN), Secondary Visual Network, and Cingulo-Opercular Network (CON). CAP-5 revealed increased activation in the auditory association cortex and MT+ complex, alongside regions associated with the Default Mode Network (DMN). CAP-7 demonstrated increased activation in the early auditory region and areas associated with the CON. These findings highlight the central role of the SMN in physics-related neural processing. The SMN appears crucial for both externally directed cognitive processes (CAP-1), which facilitate engagement with the learning environment and acquisition of new knowledge, and internally directed processes (CAP-5 and CAP-7), which support retrieval of prior knowledge and construction of new mental representations. This pattern of SMN involvement provides support for the universality of embodied cognition theory in physics learning across different teaching methodologies. The findings are in line with previous fMRI studies that have shown somatomotor recruitment during physics concept learning and have demonstrated the strong association between motion-based physics concepts (e.g., wavelength, frequency) and somatomotor activation^51,52.

Regarding changes in temporal dynamics from our pre-instruction to post-instruction contrast, a notable observation was that physics instruction led to significant changes in neural engagement, especially in the way students construct and utilize conceptual representations during problem solving. During the FCI task, we observed a notable increase in the temporal fraction of CAP-6, a pattern primarily characterized by activations in the Secondary Visual Network and the DAN, at post-instruction. Given that the FCI task requires familiarity with Newtonian principles governing objects in motion and at rest, effective conceptual reasoning relies on constructing internal schemas⁵³. These schemas allow students to form representations of potential solutions and the constraints of a problem, enabling them to quickly recognize patterns and draw upon prior knowledge to solve novel problems. During physics instruction, students encounter numerous problems related to Newtonian motion. Through repeated exposure, they develop strategies for identifying key visual features and constraints that determine an object’s trajectory. Furthermore, in contrast to novices, experts tend to spend more time constructing robust representations during the problem-solving process⁵⁴. Consequently, at post-instruction, the increased engagement of the CAP associated with activations in the Secondary Visual Network, important for object recognition and mental simulation⁵⁵, and the DAN, critical for focusing attention on relevant features⁵⁶, likely reflects students’ ability to generate more structured representations of motion problems and demonstrates expert-like thinking. This enhanced strategy supports the identification of features essential to applying Newtonian principles, thereby facilitating more effective problem-solving. Additionally, during the PK task, we observed an increase in both counts and temporal fraction within CAP-5 at post-instruction. These findings demonstrate the relevance of embodied cognition for physics-related semantic retrieval, as indicated by the involvement of the SMN. In addition, these results are corroborated by prior findings from Mason and Just⁵², which indicated that various physics concepts, such as those related to the causality of forces, flow of energy, and periodicity, were associated with activation of regions in the SMN. Additionally, this CAP showed a more refined pattern of activation that also included the MT+ Complex and Auditory Association Cortex. This CAP may be another reflection of increasing expertise as students develop more specialized neural representations that integrate visual motion processing and auditory associations, potentially drawing on visual recall of familiar diagrams or equations and auditory recall of verbal explanations^57,58, while maintaining embodied components of PK.

In addition to task-related changes, our analyses revealed that the intrinsic organization of resting-state networks was reconfigured following physics instruction, suggesting a broader pattern of neural re-configuration⁵⁹. Evidence suggests that learning interventions can induce changes in these resting-state patterns⁶⁰. In line with our hypotheses, we observed significant changes from pre- to post-instruction in two complementary brain states: increased counts, temporal fraction, and persistence of CAP-7, characterized by SMN and CON activation. This result suggests that physics learning, regardless of instructional methodology, heavily relies on embodied cognitive processes. Mason and Just⁵¹ demonstrated this relationship in their investigation of how naive participants learn physics concepts. In that study, participants viewed and learned about mechanical systems, such as a bathroom scale with its levers, springs, and mechanisms. Throughout the learning process, from initial encoding to forming mental representations and understanding system functionality, participants consistently engaged in somatomotor regions, highlighting the embodied nature of physics learning. Furthermore, the involvement of the CON may be demonstrative of the importance of sustaining attention, while simultaneously monitoring for naive or incorrect beliefs⁶¹ about foundational physics concepts. Ultimately, the changes we observed in resting state dynamics in the present study align with principles of Hebbian learning and metastability. The repeated activation of these CAPs during physics learning appears to result in internal neural restructuring, potentially optimizing the brain for future acquisition of advanced physics concepts that require sensorimotor and visuospatial processing for initial comprehension⁶². This pattern suggests that initial physics concept learning relies heavily on sensorimotor processing, and the increased engagement of this CAP during post-instruction resting state may reflect neural restructuring that facilitates sensorimotor integration in physics learning. Interestingly, in contrast to our hypotheses, we observed decreased counts, temporal fraction, and persistence of CAP-4 from pre- to post-instruction. This CAP, characterized by increased activation in the DMN, Language Network, and FPN, most likely represents more effortful processing of physics concepts. However, this decline aligns with theories of skill acquisition and automaticity development^63,64, which posit that initial learning stages require more controlled processing before transitioning to automatic processing, typically marked by reduced activation in control network regions as expertise develops. In the context of resting state activity, this reduction in CAP-4 engagement could reflect intrinsic cortical reconfiguration that primes the brain for more efficient physics concept processing⁶⁵. Similarly, we observed decreased counts and temporal fraction of CAP-1 from pre- to post-instruction, a multimodal pattern showing significant correspondence with the SMN, Visual Network, CON, and DAN. This finding for CAP-1 should be interpreted cautiously, considering that it also exhibited a significant interaction effect for resting-state persistence. Consequently, this interaction effect for persistence may be influencing the general pattern of a decline in temporal fraction and counts from pre-instruction to post-instruction due to the dependent relationship amongst these temporal metrics⁴¹. One possible explanation may be due to the strong anti-correlation between CAP-1 and CAP-4, potentially indicating these patterns may operate in tandem, with decreases in effortful processing (i.e., CAP-4) leading to subsequent decreases in CAP-1 engagement at post-instruction.

Furthermore, beyond the general post-instruction learning effects, our findings revealed that different physics instruction methods significantly influenced the temporal dynamics of physics-related brain states, with distinct patterns emerging between lecture-based and active learning methodologies. During the FCI task, students in lecture-based environments exhibited decreased CAP-4 persistence from pre- to post-instruction, whereas those in active learning environments demonstrated enhanced persistence of this pattern. A critical requirement of the FCI task is the suppression of intuitive but non-Newtonian conceptualizations of motion trajectories⁵³. CAP-4 shows its strongest positive activation in the anterior cingulate cortex (ACC), a region essential for cognitive control and particularly for suppressing competing conceptual frameworks⁶⁶. This aligns with findings from Allaire-Duquette et al.⁶⁷, which demonstrate that physics experts consistently engage neural mechanisms to suppress intuitive physics concepts when contextually required. This suggests that active learning environments may facilitate the development of more expert-like neural patterns. Moreover, the ACC not only suppresses errors but also integrates past feedback to guide future decision-making processes⁶⁸. Engagement of the posterior cingulate cortex and lateral temporal regions in CAP-4 provides additional insight into the learning process. These regions may support episodic memory integration, specifically the incorporation of experiential knowledge into conceptual understanding^69,70. This activation pattern is particularly significant from the perspective of the embodied cognition framework. Active learning environments typically incorporate direct physical engagement with physics concepts through model manipulation and kinesthetic learning experiences. Consequently, the increased CAP-4 persistence, which entails sustained periods of integrating physical classroom experiences while simultaneously suppressing non-Newtonian intuitions, may reflect the effects of embodied learning aspects often employed in active learning contexts. Additionally, concurrent activation of the Dorsolateral Prefrontal Cortex and Orbitofrontal Cortex further supports this interpretation, as these regions are instrumental in integrating multiple information streams during decision-making processes^71,72,73. This evidence suggests that the enhanced CAP-4 persistence observed in active learning students during FCI tasks reflects a sophisticated cognitive strategy that integrates embodied experience and active suppression of non-Newtonian intuitions during periods of self-referential thinking. Furthermore, prominent activation of the ACC and the auditory association areas positions the active learning group to engage in greater insight problem solving^74,75, allowing students to potentially recognize unproductive cognitive strategies and transition into more effective ones. Consequently, such a brain state may suggest that active learning students approach physics-related conceptual reasoning in a fundamentally different manner compared to their lecture-based instruction counterparts. Evidence for this distinction emerged from the main effect of class in an FCI task, where the lecture-based group exhibited greater persistence of CAP-2, a pattern characterized by activations in the Default Mode and FPNs. Although CAP-2 shares many overlapping active regions with CAP-4, it shows notably less ACC activation but substantially more activation in the inferior parietal cortex, a brain area significantly involved in visuospatial processing⁷⁶. This pattern suggests that, when solving physics conceptual reasoning problems related to Newtonian motion, active learning students may rely more on their direct physical experiences with the concepts. By contrast, students taught primarily through lectures may rely more heavily on visuospatial processing strategies. Overall, these findings imply that students’ learning environments can meaningfully influence how they engage in and approach physics conceptual reasoning tasks.

The instructional methodology differences were also evident in the PK task, where the lecture-based group exhibited increased counts of CAP-7, characterized by primary activations in both the CON and SMN, from pre- to post-instruction, when compared to the active learning group. This finding aligns with previous research demonstrating that embodied cognition plays a fundamental role in understanding physics concepts, regardless of the instructional methodology used to present the information. Notably, CAP-7 exhibits prominent activation in the Insula and Frontal Operculum, implicated in language processing and multisensory integration^77,78, and in the Early Auditory region, which is integral for processing speech⁷⁹. Additionally, previous studies have demonstrated that the CON is involved in maintaining cognitive control for both word and speech recognition^61,80. Given that lecture-based environments are typically instructor-driven and emphasize rote, recognition-based memorization⁸¹, the increased state initiation of CAP-7 may suggest that while embodied cognition may be agnostic to instructional methodology, the chosen instructional approach may induce additional neurobiological changes that mirror the manner in which the content was disseminated. Specifically, lecture-based courses may require enhanced monitoring and sustained attention to external auditory cues, promoting the formation of internal, verbally mediated representations of physics concepts. During semantic retrieval, this may necessitate internal monitoring to ensure that the recognition of physics terms and their associated meanings remains consistent with prior lecture exposure and the associated sensorimotor experiences used to understand these concepts during learning. Furthermore, the finding that lecture-based students exhibited increased counts (state initiation) of CAP-7 at post-instruction, rather than increases in persistence or temporal fraction, suggests that lecture-based instruction engages SMN-related patterns differently. While SMN activation remains essential for understanding physics concepts across instructional methods, lecture-based learning may result in more transient activation patterns. Specifically, CAP-7 may have evolved into a state that is briefly initiated to rapidly retrieve verbal representations of physics terminology from memory, rather than being stably maintained throughout the task. This transient engagement reflects how lecture-based methodology shapes the temporal dynamics of SMN activation, even though such activation remains fundamental to physics learning. Moreover, the main effect of class showed that lecture-based instruction groups displayed increased persistence of CAP-2 and CAP-3 compared to the active learning group. CAP-2 is characterized by activations in the DMN and FPN, while CAP-3 involves the DMN, FPN, and CON. Considering the high anti-correlation between CAP-2 and CAP-7, along with prominent activations in working-memory regions (Dorsolateral Prefrontal Cortex)⁷², autobiographical memory regions (Posterior Cingulate and Lateral Prefrontal Cortex)^69,70, and visual word recognition (Inferior Parietal Lobule)⁸², CAP-2 may work in tandem with CAP-7, demonstrating a toggling between internal retrieval of semantic-related memories (i.e., CAP-2) and extended periods of external focus to correctly recognize specific physics terminology (i.e., CAP-7)⁸⁰. Lastly, CAP-3 features activation in regions associated with working memory (Dorsolateral Prefrontal Cortex, Superior Parietal Lobule)⁸³, visual processing (Superior Parietal Lobule, Inferior Parietal Lobule)⁸⁴, language processing (Inferior Frontal)⁸⁵, and recognition memory (Subcortical)^86,87, which may further highlight how content is processed or encoded in lecture-based students. Considering that lecture-based instruction often emphasizes memorization and observational learning, these findings suggest a reliance on a more recognition and externally-based approach to physics-related semantic retrieval.

Furthermore, these instructional methodology differences were most prominently reflected in our resting state analyses, where, in direct support of our hypotheses, the active learning group exhibited a marked increase in the persistence of CAP-1 during resting state, from pre- to post-instruction, compared to the lecture-based instruction group. CAP-1 is a multimodal brain state characterized by activations across several networks, including the Secondary Visual Network, DAN, CON, and SMN. We posited that active learning would preferentially engage CAPs associated with sensorimotor processing, given that active learning courses often employ more kinetic, hands-on approaches, which, according to embodied cognition literature, should increase activation in sensorimotor regions^25,51. Additionally, also aligned with our hypotheses, the lecture-based instruction group demonstrated increased counts and temporal fraction of CAP-6 from pre- to post-instruction during the PK task, reflecting that observation is one of their primary modes of learning. This aligns with research showing that observational learning activates visual processing regions and the Superior Parietal Lobule^44,88. The predominance of visual instruction in lecture-based settings appears to strengthen these visual processing networks. Both of these findings suggest a functional reorganization of the brain that becomes evident even at rest. Specifically, in the active learning group, this reorganization enhances sensorimotor integration, whereas in the lecture-based group, it enhances visual activation. These findings align with the principle of Hebbian learning, where persistent activations of certain network activation patterns, in response to repeated engagement with specific tasks, lead to intrinsic cortical restructuring and stable, recurrent state^30,31,32. For both classes, such reorganization could promote improved comprehension of the physics material in the modality through which it is primarily delivered. In particular, for the active learning class, this reorganization may facilitate more efficient processing of physics concepts⁶⁵, which typically requires substantial sensorimotor engagement during the learning process⁵¹. Notably, although CAP-1 declined in both temporal fraction and counts from pre- to post-instruction among all students, its relative preservation among active learning students likely reflects the distinct demands of their learning environment. Within our study, the active learning group underwent a semester of physics instruction that employed the modeling instruction approach. This approach increases student participation through inquiry-based lab activities, collaborative group discussions, and the creation of multiple representations (i.e., graphs, diagrams, etc.) of complex physics concepts. The method establishes a multimodal learning environment that allows students to develop a more integrative and intuitive understanding of foundational physics concepts. Most importantly, increased physical engagement through lab demonstrations and drawing visual representations is expected to recruit somatomotor areas^20,89. This grounding of abstract physics concepts into concrete representations during learning may facilitate a more intuitive understanding of the material. Moreover, over time, such recurrent exposure and integration encourage long-term network reorganization that remains primed for reactivation to support future learning, potentially enabling a more cohesive comprehension of physics content or facilitating more effective absorption of material in the primary instructional mode.

Finally, despite our study revealing numerous differences in the temporal dynamics of certain CAPs differing across time and instructional methodology, these differences did not translate to observable brain-behavioral changes. Our exploratory behavioral assessment revealed only pre-instruction to post-instruction differences, with no significant differences between active and passive learning groups. This contrasts with prior studies that have established numerous behavioral advantages of active learning approaches, including improved conceptual understanding of complex physics concepts. However, this lack of behavioral differences does not diminish our neurobiological findings, as neurobiological changes do not always manifest as identifiable behavioral differences⁹⁰. Our results may represent nascent neural changes that have yet to translate into observable behavioral outcomes. Alternatively, both groups may employ different neurocognitive strategies that achieve similar behavioral results. Another possibility is that our fMRI tasks—the FCI task, which focuses on introductory Newtonian motion, and the PK task, which tests basic physics terminology—may have been sensitive enough to detect neurobiological differences between instructional groups but too simple to reveal behavioral differences once students have mastered these foundational concepts, as evidenced by the behavioral improvements observed across time. Furthermore, while our exploratory behavioral findings showed post-instruction improvements (i.e., increased FCI accuracy, increased PK accuracy, and decreased PK reaction time), our brain-behavioral analysis revealed no significant correlations. Specifically, changes in CAP-6 temporal fraction or persistence did not explain the observed behavioral improvements. This lack of brain-behavioral correlations may stem from insufficient statistical power; detecting correlations for diffuse brain states requires larger sample sizes than localized voxel-wise analyses, with even 150 participants potentially being inadequate⁹¹. Additionally, our linear analytical approach may have failed to capture the relationship, as brain-behavioral associations could be more complex and non-linear.

The current study has several limitations. First, our original sample size was reduced from 121 to 90 participants due to our stringent exclusion criteria, which required at least one functional run of both resting state and task-based fMRI data for both pre- and post-instruction sessions; participants were excluded for various reasons including non-return for post-instruction scans, technical issues, MRIQC flagging for high mean framewise displacement (FD), ghost-to-signal ratio, low signal-to-noise ratio, entropy focus criterion, or missing functional runs from one or more tasks. This criterion was applied to ensure CAPs were not confounded by functional runs flagged due to poor quality control and to minimize interpretational ambiguity that could arise when comparing CAPs across different conditions. Specifically, by requiring participants to contribute data across both sessions and multiple tasks, we reduced the risk that observed differences in temporal dynamics would be attributable to different participant compositions rather than to the inherent cognitive demands of the tasks themselves. Importantly, t-tests and chi-square tests revealed no demographic differences between the active learning and lecture-based groups, and the final sample sizes were similar, with 46 active learning and 44 lecture-based participants. Furthermore, while the optimal sample size depends on the study and the specific effect being investigated, neuroimaging research can achieve adequate statistical power even with relatively small samples. Moreover, when averaging across subjects, the central tendency of the brain’s functional organization can be accurately estimated in a small sample size consisting of 25 participants (the median sample size for neuroimaging studies)⁹², which the sample size used in our study, consisting of 90 participants, exceeds. This generalizability is further enhanced when assessing large areas of activation⁹³, which pertains to our study, which examined CAPs across large-scale networks. Second, our analytic approach identified shared CAPs across resting state and task-based fMRI data, which facilitated comparisons across instructional methods but did not allow for the detection of potential group-specific CAPs that might have emerged if clustering had been performed separately. Our chosen approach aimed to improve the interpretability and comparability of results across different instructional methods. However, this methodological decision means that potential group-specific CAPs, which might be preferentially influenced by a particular instructional methodology, were not identified. Consequently, the study did not examine the temporal dynamics of these group-specific CAPs, which could have provided deeper insights into how different instructional approaches uniquely shape distinct patterns of network activations. Furthermore, within this study, a combined clustering approach was used to identify generalized CAPs related to multiple aspects of physics cognition (i.e., conceptual reasoning and knowledge retrieval) while assessing how these CAPs are engaged during physics-related tasks and rest. Future studies may consider separate clustering approaches for each instructional group or task to better explore unique neural activation patterns. Third, the exclusion of volumes with FD exceeding 0.35 mm was necessary to prevent high-motion artifacts from confounding the clustering results and ensure CAP integrity⁹⁴; however, this may have impacted temporal continuity metrics such as persistence. This approach was essential to prevent high-motion volumes from confounding the clustering results, as including frames with excessive motion would standardize volumes with minimal artifacts against those with significant head motion, potentially distorting the global mean and leading to misleading CAP identification. Additionally, frames predominantly influenced by head motion rather than true neuronal activity might be erroneously clustered together, generating CAPs driven more by motion artifacts than by meaningful brain states. Furthermore, the participants included in the analysis did not have more than 40% of their frames scrubbed from their resting state data and 20% of their frames scrubbed from the physics-only condition of their task data (both FCI and PK), aligning with prior CAPs studies that applied scrubbing, retained participants with less than 30-50% of their frames scrubbed, and still computed temporal continuity metrics equivalent to persistence. Additionally, as shown in Supplementary Table 4, the average length of consecutive scrubbed segments across each task was relatively low, demonstrating that predominantly short, isolated segments were removed rather than long contiguous portions of the time series, thereby minimizing the impact on temporal continuity. Moreover, consistent with prior studies that scrubbed volumes due to excessive motion in preparation for CAPs analysis, we prioritized original data integrity for the final analysis by leveraging the remaining data as opposed to interpolating missing volumes, which would then be treated as observed brain patterns in the analysis^95,96,97. Finally, the current study was limited to examining pre- to post-instruction changes in temporal dynamics among introductory physics students over a single semester. While previous evidence has shown that instruction-related neurobiological changes can occur within a single semester⁶⁰ and changes in dFC can manifest shortly after an intervention^98,99, our study is limited in the long-term implications of our findings. The neurobiological differences that we observed between modeling and lecture-based students may result in differences in how students engage with, comprehend, and build upon physics concepts in their subsequent courses, and potentially shape their longer-term cognitive abilities related to scientific reasoning. Consequently, future studies should aim to examine how the neurobiological differences between instructional approaches evolve and persist over a longer time frame to understand the longer-term consequences associated with these different instructional methodologies.

Overall, in this study, we examined how university physics learning acquired via different instructional methodologies (i.e., active versus passive learning) is associated with varying temporal dynamics of brain states. Our findings indicate significant intrinsic neural restructuring over a semester of learning. In line with our hypotheses, we found that physics learning induced changes across multiple networks (e.g., frontoparietal, somatomotor, and visual processing networks). Specifically, active learning students showed increased persistence of an integrated pattern with elevated SMN activation (CAP-1) at post-instruction. This finding suggests that the collaborative, multimodal, and hands-on approaches of active learning preferentially engage integrated SMN patterns to ground physics concepts, supporting an embodied cognition framework. In contrast, lecture-based students exhibited greater persistence of a visuospatial network (CAP-6), suggesting a reliance on visual observation strategies. Moreover, additional evidence demonstrating the universality of embodied cognition in physics learning was exhibited by the lecture-based group’s increased counts (state initiation) of CAP-7, a pattern characterized by SMN and CON activation, during physics semantic retrieval. This finding suggests that while SMN engagement remains fundamental to physics understanding regardless of instructional approach, the temporal dynamics of SMN-related patterns differ between methodologies. Specifically, lecture-based instruction may promote transient activation rather than sustained engagement of SMN-related patterns, highlighting how instructional methods shape the temporal characteristics of embodied cognitive processes. Taken together, these findings provide evidence that the kinetic, hands-on approaches employed in active learning classrooms yield measurable differences in brain activation compared to traditional lecture-based instruction. These results enhance our understanding of how university physics education influences brain dynamics. Furthermore, this work indicates that instructors’ pedagogical choices influence the neural strategies that students employ during physics cognition, ultimately playing a role in reshaping the neural architecture of learning. Finally, while prior literature suggests that the embodied cognition framework extends beyond physics education (e.g., gesture use in mathematics learning)¹⁰⁰, the current study demonstrates its neurobiological relevance specifically within introductory physics education. By comparing two instructional approaches, including modeling instruction that is uniquely tailored to physics education, we provide neural evidence for embodied cognition in this domain. Future studies should investigate the generalizability and neurobiological relevance of this framework when comparing active versus lecture-based instruction across other science, technology, engineering, and mathematics disciplines.

Methods

Participants

The sample consisted of 121 healthy, right-handed, undergraduate students (mean age = 19.8 ± 1.5, range = 18–26 years; 55 females) enrolled in a 15-week semester of calculus-based, introductory physics course at FIU. Students received either traditional, lecture-based instruction (n = 60) or modeling instruction (n = 61). Modeling instruction is a student-centered active learning approach to physics education where learners actively construct, test, and refine conceptual models through collaborative activities. Students engage in laboratory demonstrations and peer-led discussions with minimal instructor intervention, with the instructor only interjecting to prompt deeper conversations or clarify concepts as needed. In both small groups and whole-class settings, students communicate their understanding with their peers using multiple representational tools, such as graphs, equations, and diagrams. This methodology is designed to foster a deeper conceptual understanding of physics principles through active engagement rather than reception of information^45,101. At enrollment, participants provided demographic information, including age, sex, ethnicity, household income, GPA, and years enrolled at FIU (i.e., freshman, sophomore, junior, or senior) (Table 1) (Supplementary Table 1). All participants were free from cognitive, neurological, or psychiatric conditions and reported no use of psychotropic medications.

Table 1 Participant demographic information

Procedures

Behavioral and neuroimaging data collection occurred during two time periods: pre- and post-instruction. Multiple different recruitment strategies were employed, including several in-class visits (at the professor’s discretion), emailing eligible students, and distributing study flyers across FIU’s campus. Enrolled participants completed an fMRI session at the start of the course (i.e., pre-instruction), no later than the fourth week of instruction, and before the first exam. Neuroimaging sessions were conducted off-campus, and participants were provided with free parking and/or FIU-organized transportation to and from the MRI site. At the end of the semester, participants completed their post-instruction data collection session no later than 2 weeks after the semester concluded and after final exams were completed. Written informed consent was obtained in accordance with FIU’s Institutional Review Board approval (14--0007). Additionally, participants were monetarily compensated for their study participation.

MRI data acquisition

MRI data were acquired on a GE 3 T Healthcare Discovery 750 W MRI scanner at the University of Miami. Functional imaging data were acquired with an interleaved gradient-echo, echo planar imaging sequence (repetition time [TR]/echo time [TE] = 2000/30 ms, flip angle = 75°, field of view (FOV) = 220 × 220 mm, matrix size = 64 × 64, voxel dimensions = 3.4 × 3.4 × 3.4 mm, 42 axial oblique slices). T1-weighted structural data were also acquired using a 3D fast spoiled gradient recall brain volume (FSPGR BRAVO) sequence with 186 contiguous sagittal slices (TI = 650 ms, bandwidth = 25.0 kHz, flip angle = 12°, FOV = 256 × 256 mm, and slice thickness = 1.0 mm).

fMRI tasks

Participants completed a self-paced, conceptual reasoning task consisting of questions derived from the FCI (Fig. 6A–C)^102,103,104 that was adapted for the MRI environment⁵³. FCI trials included textual descriptions and illustrations of scenarios, including objects at rest or in motion, accompanied by one correct Newtonian solution and several incorrect non-Newtonian solutions. Control trials included similar visual stimuli to FCI trials but tested general reading comprehension and/or shape discrimination. Each trial was presented in blocks composed of three sequential view screens: Phase I: Scenario, which included an illustration and text describing a physical scenario; Phase II: Question, which involved presenting a question related to the scenario; and Phase II: Answer, where participants were presented with four answer choices. Both FCI and control blocks had a maximum duration of 45 s and were followed by a fixation cross with a minimum duration of 10 s. The total duration for each FCI run was 5 min 44 s; data were collected during three runs for a total duration of ~16 min, resulting in 170 volumes being collected per participant¹⁰⁵. Moreover, to ensure that we were examining the understanding of Newtonian principles as opposed to familiarity with specific questions, different sets of questions for the physics and non-physics trials were presented for the pre- and post-instruction fMRI sessions.

Fig. 6: Force Concept Inventory (FCI) and physics knowledge (PK) tasks.

Example items of the in-scanner physics tasks, including the three phases of the FCI task A Phase I: Scenario, B Phase II: Question, C Phase III: Answer, and D the PK task. Reprinted from Smith et al.¹⁰⁵, Task-based attentional and default mode connectivity associated with science and math anxiety profiles among university physics students, Pages 100204, Copyright (2023), with permission from Elsevier.

Participants also completed the PK task (Fig. 6D), which was adapted from a general knowledge task of semantic retrieval¹⁰⁶. The PK task was presented in a block design and probed for brain activation associated with physics-based content knowledge. Students viewed physics questions related to general PK and their corresponding answer choices. A control condition was presented in which students viewed general knowledge questions with corresponding answer choices. PK and control blocks were 28 s long and included four questions per block (6.5 s per question followed by 0.5 s of quick fixation). Three blocks of physics or general questions (six question blocks total) were alternated with 10 s of fixation. The total duration of one run was 4 min 2 s; data were collected during two runs for a total duration of ~8 min, resulting in 180 volumes being collected per participant⁹⁵. Furthermore, similar to the FCI task, different sets of physics and control questions were utilized for the pre- and post-instruction fMRI sessions.

Participants were instructed to lie quietly with their eyes closed and not to fall asleep. The total duration of the resting state session was 12 min, resulting in 365 volumes of data being collected per participant. The first five volumes for each participant were removed to ensure only steady-state data were used in subsequent analyses, leaving 360 volumes of resting-state data per participant.

MRI preprocessing

As stated in Smith et al.¹⁰⁵, all functional runs within this study (i.e., FCI, PK, rest) were preprocessed independently. Each participant’s T1-weighted images were corrected for intensity non-uniformity with ANT's N4BiasFieldCorrection tool^107,108. Anatomical and functional images underwent additional preprocessing steps using fMRIPrep (v.1.5.0rc1), a BIDS-compliant Python package that automatically provides a high-quality, reproducible preprocessing workflow with minimal user intervention¹⁰⁹. The T1-weighted (T1w) reference, which was used throughout the pipeline, was generated after T1w images were corrected for intensity non-uniformity with ANT's N4BiasFieldCorrection. Freesurfer's mri_robust_template was used to generate a T1w reference, which was used throughout the entire pipeline^108,110. Nipype's implementation of ANT's antsBrainExtraction workflow was used to skullstrip the T1w reference using OASIS30ANTs as the target template¹¹¹. FSL's FAST was used for brain tissue segmentation of the cerebrospinal fluid (CSF), white matter (WM), and gray matter; brain surfaces were reconstructed using Freesurfer's recon_all^112,113. All anatomical images were spatially normalized to the MNI152NLin2009cAsym template, specifically using non-linear registration with antsRegistration (ANTs 2.2.0)¹¹⁴.

Preprocessing of functional images began with selecting a reference volume and generating a skull-stripped version using a custom methodology of fMRIPrep. Susceptibility distortion correction employed a fieldmap-less method where deformation fields were derived by registering the BOLD reference to the intensity-inverted T1w reference¹¹⁵. AntsRegistration (ANTs 2.2.0) performed this registration with deformation constraints limited to the phase-encoding direction and regularized using an averaged fieldmap template¹¹⁶. The resulting distortion estimates generated an unwarped BOLD reference to enhance anatomical co-registration precision. Freesurfer's bbregister, which uses boundary-based registration, was used to coregister the T1w reference to the BOLD reference. The BOLD time series were then resampled onto surfaces of fsaverage5 space, resampled into standard MNI152NLin2009cAsym space, and resampled onto their original, native space by applying a single, composite transform to correct for head motion, using mcflirt (FSL 5.0.9)¹¹⁰, and susceptibility distortions. In the pipeline, antsApplyTransforms (ANTs) with Lanczos interpolation was used to perform gridded (volumetric) resamplings, and mri_vol2surf (FreeSurfer) was used to perform non-gridded (surface) resamplings. Additionally, the BOLD time series were high-pass filtered, using a discrete cosine filter with a cutoff of 128s¹¹⁷. For each volume, except the first, Nipype was used to compute FD (as described in Power et al.¹¹⁸). Nuisance signals from CSF, WM, and whole brain masks were extracted by using a set of physiological regressors, which were extracted to allow for both temporal component-based noise correction (tCompCor), which used the most variable 5% of voxels within an eroded subcortical brain mask, and anatomical component-based noise correction (aCompCor), which used the used the intersection of the subcortical mask and the combined CSF/WM masks projected from T1w to native functional space¹¹⁹. Additionally, the confound time series derived from head motion estimates were expanded to include their temporal derivatives and quadratic terms, resulting in a total of 24 head motion parameters (i.e., six base motion parameters, six temporal derivatives of six motion parameters, 12 quadratic terms of six motion parameters, and their six temporal derivatives). Estimates for the global, CSF, and WM signals were expanded to include their temporal derivatives and quadratic terms, resulting in a total of 12 signal-based parameters (i.e., three base signal parameters, three temporal derivatives of the three base parameters, the three quadratic terms of the base parameters, and the three quadratic terms of the temporal derivatives). Finally, all 24 head motion confound estimates, three high-pass filter estimates, and a variable number of aCompCor estimates (components that explain the top 50% of the variance) were outputted into a TSV file to be used for later denoising steps¹²⁰.

Quality control assessment

Prior to time series extraction, MRIQC¹²¹ was used to calculate quality metrics for each participant’s functional runs, including mean FD, ghost-to-signal ratio, signal-to-noise ratio, and entropy focus criterion (an indicator of ghosting and blurring resulting from head motion)¹²². Based on the MRIQC reports, participants’ functional runs with a mean FD, ghost-to-signal ratio, or entropy focus criterion above the 99th percentile, or a signal-to-noise ratio below the 1st percentile, were excluded from the analysis. This quality control procedure resulted in the exclusion of all functional runs from both pre- and post-instruction sessions for two participants in the resting state data, two participants in the FCI task, and one participant in the PK task. Moreover, the participants excluded from the FCI and PK tasks overlapped with those excluded from the resting state. Additionally, twelve participants had all functional runs excluded for either their pre-instruction (n = 6) or post-instruction (n = 6) resting state session, where only one run per session was collected. In total, MRIQC flagged 14 unique participants who had incomplete datasets for at least one task (resting state, FCI, or PK) due to high-motion artifacts (including blurring and ghosting) and low signal-to-noise ratio (e.g., thermal noise).

Denoising and parcellation

Following quality control assessment, all runs and sessions from both the resting state and task data underwent identical denoising strategies. This uniformity was essential to ensure that the subsequent clustering of time points from the combined dataset was not influenced by variations in denoising approaches. Specifically, the time series of each participant’s task and resting state data were bandpass filtered (0.008–0.1 Hz), detrended, and subjected to nuisance regression. This regression involved orthogonalizing the time series against the six head motion parameters and their first derivatives, the global signal and its first derivative, and the first six aCompCor components derived from the WM and CSF masks. Spatial dimensionality reduction was performed using the Human Connectome Project extended (HCPex) atlas, which is both a volumetric version and extension of Glasser et al.⁴⁷ that incorporates an additional 66 subcortical areas, resulting in a total of 426 nodes⁴⁶. Additionally, Huang et al.⁴⁶ classified these 426 nodes into 23 cortical and subcortical divisions (Fig. 7).

Fig. 7: Cortical and subcortical divisions of the HCPex Atlas.

The HCPex Atlas, which contains 426 nodes, was used to parcellate each participant’s preprocessed fMRI data. All 426 nodes within the atlas were classified into 23 cortical subdivisions, as described by Huang et al.⁴⁶.

For task data, only volumes corresponding to the physics trials were retained (i.e., control trials were removed), resulting in ~40 and 45 frames retained per run for the FCI and PK tasks, respectively. To account for the hemodynamic delay¹²³, onset times for each physics trial block were shifted forward by 4 s¹²⁴. Additionally, for both the physics task and resting state data, volumes from each functional run with an FD exceeding 0.35 mm were scrubbed. These TRs were removed to prevent clustering from being significantly affected by time points dominated by motion artifacts. Participants were also assessed to ensure that the proportion of scrubbed volumes did not exceed 20% per run for the physics task data and 40% per run for the resting state data. A stricter retention threshold was applied to the task data due to the difference in the number of frames being assessed (i.e., 40 and 45 frames per run for the FCI and PK tasks, respectively), compared to 360 frames for resting state. These retention thresholds align with previous, recent CAP studies that retained participants with less than 30-50% of scrubbed data^95,96,97. Afterwards, standardization was conducted within runs to allow the subsequent clustering procedure to prioritize patterns of network activation in all HCPex regions of interest (ROI) as opposed to the absolute magnitude differences, especially between task and rest data.

Coactivation pattern (CAP) analysis

Once denoising was completed, CAP analysis was performed. To this end, participants with at least one functional run across all sessions and tasks were concatenated into a single matrix (rows: ${N}_{{participants}}$*${N}_{{Repetition\; Time}}$; columns: ${N}_{{ROIs}}$). This resulted in the exclusion of 31 participants: 14 where all functional runs for a specific session and task were flagged by MRIQC for quality control issues (excessive motion-related artifacts and low signal-to-noise ratio), and 17 with incomplete datasets due to factors, such as not returning for the post-instruction scanning session, resulting in a final sample size of 90 participants. Moreover, this criterion was applied to mitigate bias in clustering solutions and subsequent statistical analyses that could arise from the overrepresentation of data from specific sessions or tasks, or from participants of certain demographics contributing data exclusively to particular tasks and sessions. This approach allows for better identification of CAPs and statistical trends that are more broadly generalizable across all tasks and sessions. The matrix was then subjected to k-means clustering using Euclidean distance, where values of $k$ ranging from 2 to 20 were tested. The optimal value of cluster size was determined using the elbow criterion, a heuristic where k is the value where the within-cluster sum of squares (i.e., inertia) stops decreasing exponentially¹²⁵.

After the optimal value of $k$ was determined, the cluster centroids, representing the average of all data points assigned to the same CAP, were extracted to generate five visualizations: a regional heatmap, a correlation matrix, surface plots for each CAP, and radar plots for each CAP. The regional heatmap was created by extracting each cluster centroid, calculating the average activation among all nodes assigned to the same cortical division within the HCPex atlas, as described by Huang et al.⁴⁶, and visualizing these averages. The correlation matrix was obtained by calculating the Pearson correlation between all possible pairs of CAPs, with p values generated for each element in the matrix, to assess the linear relationship between each pair of CAPs. Surface plots were created by replacing each label in the HCPex atlas with its corresponding value from the cluster centroid and converting the atlas from MNI volumetric space into FreesurferLR (fsLR) surface space for visualization. Radar plots for each CAP were constructed by generating separate activation vectors for the “High Amplitude” activations, which are values above zero, and the “Low Amplitude” activations, which are values below zero. The cosine similarity between each activation vector and a binary representation of each a priori cortical division in the HCPex atlas^97,126,127, was computed as shown in Eq. (1):

$${co}{s(\theta )}_{A,\,C}=\frac{{cap}\_{centroi}{d}_{A}\cdot {binary}\_{vecto}{r}_{C}}{\mathrm{||}{cap}\_{centroi}{d}_{A}\mathrm{||||}{binary}\_{vecto}{r}_{C}\mathrm{||}}$$

(1)

where A ∈ {High Amplitude, Low Amplitude} and C ∈ {Cortical Division₁, Cortical Division₂, …}.

Each element within the CAP cluster centroid represented the relative activation or deactivation of an ROI belonging to one HCPex cortical division. Hence, the binary vectors representing the cortical divisions consisted of “1”s for the ROIs within the cluster centroid assigned to the same cortical division, while “0”s indicated otherwise. Furthermore, the absolute values of the negative activations within the cluster centroid were computed to restrict the “High Amplitude” and “Low Amplitude” cosine similarities from 0 to 1 for visualization purposes. Cosine similarities greater than 0.2 or less than −0.2 are reported in the results section.

Next, a network correspondence analysis was performed using the cbig_network_correspondence Python package⁴⁹. This analysis evaluated the overlap between positive activations of each CAP (with cosine similarity ≥0.20) and the Cole-Anticevic atlas, which maps ROIs from the Glasser et al.⁴⁷ atlas onto 12 distinct resting state networks⁴⁸. The package employs spin permutation methods⁵⁰ to assess overlap significance, quantified through the Dice coefficient. Finally, the CAP assignments of each participant’s time points for all runs and sessions of their task and resting state data were extracted to calculate three CAP measures for each session: temporal fraction (i.e., the proportion of time occupied by a single CAP), persistence (i.e., the average duration a CAP persists before transitioning to another CAP), and counts (i.e., the frequency of CAP initiation across the entire scan)⁴¹. Furthermore, to control the number of statistical comparisons for each task (specifically the FCI and PK tasks), multiple runs were treated as a single, continuous run for the calculation of each CAP metric¹²⁸.

Statistical analyses

We first assessed demographic differences between active learning and lecture-based instruction groups. Independent samples t-tests were used to compare age and GPA (both pre- and post-instruction), while chi-square tests were used to compare sex, ethnicity, household income, and years enrolled at FIU. These analyses were conducted for both the full sample (n = 121) and the subsample of participants who had at least one functional run for all tasks and sessions (n = 90). Next, using the calculated CAP metrics (i.e., temporal fraction, counts, and persistence), we assessed the main effects of time (i.e., pre- versus post-instruction), the main effects of instructional methodology (i.e., modeling versus lecture-based instruction), and the interaction effects between time and instructional methodology. Linear mixed-effects models were employed using the lme4 package¹²⁹, regressing each CAP metric on the interaction between time and instructional methodology, while controlling for nuisance variables such as age, sex, ethnicity, household income, GPA, and number of years enrolled at FIU, and accounting for random effects due to individual differences. Furthermore, counts are raw frequency calculations of CAP initiations, unlike temporal fraction and persistence, which include some level of accounting for differences in the magnitude of TRs due to being proportion-based and average duration-based metrics, respectively. Therefore, the total number of TRs per participant for the given task being analyzed was included as a covariate in all linear mixed-effects models assessing counts. This was done to ensure that observed differences in the counts of any CAP could be attributed to effects of interest as opposed to differences in the total number of TRs (as a consequence of scrubbing or the self-paced nature of the FCI task). Further descriptive information regarding the total number of TRs for contrasts of interest can be found in Supplementary Information. Moreover, all linear mixed-effects models were probed for the main effects of time and instructional methodology using R's emmeans¹³⁰ as the beta coefficients for time and class represent mean differences from the reference group, rather than capturing differences among the variables themselves. Finally, we performed additional exploratory analyses, examining brain-behavioral correlations for findings associated with the FCI (i.e., FCI task accuracy and reaction time) and PK (i.e., PK accuracy and reaction time) tasks. First, linear mixed-effects models were created to assess the main effects of time, instructional methodology, and their interaction on the four behavioral measures (i.e., FCI task accuracy, FCI task reaction time, PK task accuracy, and PK reaction time). These models accounted for the same nuisance variables (i.e., age, sex, ethnicity, household income, GPA, and number of years enrolled at FIU) and random effects due to individual differences as the primary models assessing temporal dynamics. Second, for any contrast and behavioral measure that exhibited significant findings (from the analyses described above), we examined the relationship between these behavioral measures and significant temporal dynamic findings (identified from the CAP metric analyses) associated with the same task paradigm and contrast. Specifically, we examined the interaction between the contrast (e.g., time, instructional methodology) and the significant temporal dynamic metrics to assess their effect on the behavioral measures, using linear mixed-effects models that accounted for nuisance variables (as previously detailed) and random effects for individual differences. To account for multiple comparisons for non-exploratory analyses, the Benjamini–Hochberg procedure (BH) was applied, and findings with corrected and uncorrected p values are reported and discussed.