Machine Learning Applications in Materials Design

Machine learning has transformed the way novel materials are discovered and engineered, shifting the field from trial-and-error experimentation towards data-driven and predictive strategies. By training algorithms on large datasets of computed and experimental properties, researchers can build surrogate models that approximate complex quantum-mechanical simulations at a fraction of the computational cost. These models enable rapid screening of vast chemical spaces, identification of promising candidates for energy storage, electronics and structural applications, and expansion of known stable materials far beyond human intuition. Key approaches include graph-based neural networks that capture atomic connectivity, fragment-based descriptors that encode local environments, active-learning schemes that iteratively guide experiments or calculations, and uncertainty quantification to ensure reliable predictions. Integration with high-throughput computation and automated synthesis platforms has accelerated workflows, while interpretability methods seek to connect learned patterns with established physical principles. Together, these advances promise to shorten development cycles for advanced batteries, catalysts, photonic devices and structural alloys, and to address global challenges in clean energy, sustainability and advanced manufacturing.

Research from Nature Portfolio

Recent studies have demonstrated that graph-based deep-learning models, when trained at scale on tens of thousands of crystal structures, can predict formation energies, ionic conductivities and interatomic forces with unprecedented accuracy. Such models have enlarged the catalog of potentially stable inorganic compounds by over an order of magnitude, uncovering millions of new structures that lie below the thermodynamic convex hull and many of which have since been realised experimentally. Parallel work has introduced universal fragment descriptors that transform crystal structures into sets of property-labelled building blocks, enabling rapid and accurate prediction of band gaps, elastic moduli and thermal properties across diverse chemistries. An earlier proof-of-concept in high-throughput screening showed that machine-learning models built on chemo-structural fingerprints and electronic charge densities can deliver ultra-fast yet robust property predictions for model chain systems, illustrating how statistical learning can augment quantum-mechanical databases and guide exploratory synthesis.

Machine Learning Applications in Materials Design publication trend

The graph below shows the total number of articles in machine learning applications in materials design across all publications each year (not limited to Nature Index journals).

Technical terms

Graph neural network: A machine-learning model that represents materials as graphs, with atoms as nodes and bonds or neighbour relationships as edges, to predict properties directly from structure.

Descriptor: A numerical representation or fingerprint of a material’s composition, structure or electronic environment used as input to machine-learning algorithms.

Surrogate model: A computationally efficient machine-learning model trained to approximate results of expensive simulations or experiments.

Active learning: An iterative process in which the model identifies the most informative new data points to be measured or computed, thereby improving its predictive accuracy with minimal additional data.

Convex hull: The set of thermodynamically stable compositions in a multi-component phase diagram; structures below this hull are predicted to be metastable or stable against decomposition.

References

1. Scaling deep learning for materials discovery. Nature (2023).
2. Universal fragment descriptors for predicting properties of inorganic crystals. Nature Communications (2017).
3. Accelerating materials property predictions using machine learning. Scientific Reports (2013).
4. Accelerated search for materials with targeted properties by adaptive design. Nature Communications (2016).
5. Recent advances and applications of machine learning in solid-state materials science. npj Computational Materials (2019).
6. A strategy to apply machine learning to small datasets in materials science. npj Computational Materials (2018).