Search

Long-range interactions are challenging for machine learning interatomic potentials (MLIPs). Here, authors show that, by just learning from energies and forces, MLIPs can accurately capture electrostatics and predict atomic charges.

Accurately solving the Schrödinger equation is challenging. Here, authors present QiankunNet, a Transformer-based framework that efficiently captures quantum correlations, achieving high accuracy in complex molecular systems using neural network quantum states.

Predicting material synthesizability remains a major challenge in materials science. Here, the authors develop CSLLMs achieving 98.6% accuracy in predicting synthesizability while suggesting synthetic methods and precursors with >90% accuracy.

The dominant hydrated proton solvation structure remains debated due to ambiguous 1750 cm⁻¹ IR band assignment. Here, authors assign this signature by unveiling a distinct Intermediate configuration with the devised FFT-based vibrational analysis.

Efficient methods to calculate magnetically induced currents in metallic nanostructures are currently lacking. Here, the authors propose a theoretical method to compute and analyze magnetically induced currents in nanostructures validated for experimentally synthesized gold-based, hydrogen-containing ligand-protected clusters.

The electronic structure of benzene has been a test bed for competing theories along the years. Here the authors show via quantum chemistry calculations that the wavefunction of benzene can be partitioned into tiles which show that the two electron spins exhibit staggered Kekulé structures.

Chemical reaction networks often have multiple energy transduction pathways. Here, authors show how these pathways can be conceptualized as chemical gears and determine the most efficient operation of a network under any condition.

Machine learning prediction of electronic structure remains a general challenge. Here, the authors have developed a model with high accuracy and transferability by treating the electron density as a 3D greyscale image and adapting techniques from the image super-resolution computer vision field.

Predicting chemo- and regioselectivity is crucial for understanding and designing chemical reactions. Here, the authors introduce the projection of orbital coefficient vector (POCV) method to predict π electron properties and reactivity vectors

FIORA, an advanced graph neural network, enhances the simulation of tandem mass spectra by learning molecular bond-breaking patterns. The open-source algorithm generates high-quality reference spectra for compound identification in metabolomics.

Proton-coupled electron transfer reactions are key steps in electrocatalysis. Here the authors introduce a computational method to model such processes including nuclear quantum effects under a grand canonical constant electric potential.

Porous materials are difficult to design due to their high structural complexity. Here the authors introduce a diffusion model using signed distance functions that allows inverse design of metal-organic frameworks targeting diverse property modalities.

The accuracy of machine learning potential (MLP) is limited by the underlying ab initio methods. Here, authors refine the MLP by learning from experimental dynamical data (like spectroscopy) using an automatic differentiation technique.

The wave function contains all the information about the electrons in chemical systems. Here, authors interpolate this quantum variable through chemical interactions, allowing us to follow the electronic state as atoms move and react.

The first step in predicting material properties is the generation of a plausible crystal structure. Here, the authors introduce a large language model that can achieve this task given only the chemical composition of the material.

This Perspective highlights the potential integrations of large language models (LLMs) in chemical research and provides guidance on the effective use of LLMs as research partners, noting the ethical and performance-based challenges that must be addressed moving forward.

Modelling reactions in solution is challenging. Machine learning potentials offer promising alternatives but need large datasets. Here the authors report an automated active learning approach using descriptor-based selectors to model Diels-Alder reactions.

Predictive machine learning models, while powerful, are often seen as black boxes. Here, the authors introduce a thermodynamics-inspired approach for generating rationale behind their explanations across diverse domains based on the proposed concept of interpretation entropy.

Enhancing retrosynthetic efficiency requires overcoming the vast complexity of chemical space, the limited known interconversions between molecules, and the challenges posed by limited experimental datasets. Here, the authors introduce generative machine learning methods for retrosynthetic planning that generate reaction templates.

Structure relaxation is an iterative process particularly demanding in large-scale material discovery campaigns. Here, the authors realize a deep generative model able to relax material structures in a single step while estimating its accuracy.

An outstanding limitation of molecular dynamics simulations is sampling of long timescales. Here, authors combine Metadynamics, a popular enhanced sampling method, with stochastic resetting, to achieve higher speedups and improved kinetic inference.

Equivariant neural networks are state-of-the-art for machine learning-driven molecular dynamics (MD) simulations but have high computational cost. Here, the authors develop a Euclidean transformer that balances accuracy, stability, and speed, enabling stable long-timescale simulations of complex molecules

Stereoselective catalysts impact polymer’s properties, but discovering such catalysts is expensive and based on trial-and-error. Here, the authors develop a machine-learning tool to guide catalyst discovery and reveal mechanistic features affecting stereoselectivity.

Retrosynthesis prediction is a fundamental problem in organic synthesis. Here, inspired by simplified arrow-pushing reaction mechanisms, the authors develop a graph-to-edits framework, Graph2Edits, based on graph neural network for retrosynthesis prediction.

Predictive modelling remains a key challenge for designing synthetic transformations. Here, the authors develop a knowledge-based graph model to predict reaction yield and stereoselectivity, offering an extrapolative and interpretable approach for evaluating reaction performance.

A method is developed for the directional optimization of multiple properties without prior knowledge on their nature. Using a large ligand dataset, diverse metal complexes are found along the Pareto front of vast chemical spaces.

Neural wavefunctions have become a highly accurate approach to solve the Schrödinger equation. Here, the authors propose an approach to optimize for a generalized wavefunction across compounds, which can help developing a foundation wavefunction model.

Calculations of relative binding free energy are crucial for lead optimization in structure-based drug design, but classical methods are computationally expensive. Here, the authors describe a more efficient method for calculating the free energy that is as accurate as thermodynamic integration.

Conventional ab initio calculations and machine learning provide limited information on catalytic activity and selectivity and often show discrepancy with experimental results. Here, the authors report a high-throughput virtual screening strategy to identify active and selective catalysts, leading to the discovery of Cu-Ga and Cu-Pd catalysts for CO2 electroreduction.

Electronic structure methods are vital, yet they are often too computationally expensive. Here, the authors develop machine learned density matrices to fully represent electronic structures in a computationally cheap and accurate way.

Reticular frameworks are crystalline porous materials with desirable properties such as gas separation, but their large design space presents a challenge. An automated nanoporous materials discovery platform powered by a supramolecular variational autoencoder can efficiently explore this space.

Exciton delocalization in molecular aggregates is suggested to counteract the Energy Gap Law. Here, authors reveal the underlying physical picture and find the optimal excitonic coupling that minimizes nonradiative decay by nearly exact simulations.

Wavepacket dynamics around conical intersections are influenced by geometric phase, which can affect chemical reaction outcomes but has only been observed through indirect signatures. Now, by engineering a controllable conical intersection in a trapped-ion quantum simulator, the destructive wavepacket interference caused by a geometric phase has been observed.

Singlet oxygen relaxes in water to its triplet ground state. Here, the authors show that heavy-atom tunnelling speeds up this reaction from time scales longer than the age of the universe to microseconds.

Obtaining good initial structures is the main challenge for the computational study of transition states. Here, fast and accurate predictions for transition state of gas phase reactions are achieved by machine learning based on interatomic distances.

Deep neural networks can learn and represent nearly exact electronic ground states. Here, the authors advance this approach to excited states, achieving high accuracy across a range of atoms and molecules, opening up the possibility to model many excited-state processes.

Large language models have recently emerged with extraordinary capabilities, and these methods can be applied to model other kinds of sequence, such as string representations of molecules. Ross and colleagues have created a transformer-based model, trained on a large dataset of molecules, which provides good results on property prediction tasks.

Density functional theory provides a formal map from the electron density to all observables of interest of a many-body system; however, maps for electronic excited states are unknown. Here, the authors demonstrate a data-driven machine learning approach for constructing multistate functionals.

Retrosynthesis is a critical task for organic chemistry with numerous industrial applications. Here, the authors build a machine learning model to learn the concept of substructures from a large reaction dataset to achieve chemist-like intuitions.

Transition state energy correlations are key to the computational search for new catalysts, but are computationally expensive. Here the authors generalize a recent approach based on bond-order conservation arguments and apply it to dehydrogenation reactions on low index metal surfaces

Identifying pathways and transition states is critical to understanding chemical and biological reactions. Here, the authors introduce a capable computational approach using conformational space annealing to find multiple reaction pathways via global optimization of the Onsager-Machlup action.

Generative models for the novo molecular design attract enormous interest for exploring the chemical space. Here the authors investigate the application of chemical language models to challenging modeling tasks demonstrating their capability of learning complex molecular distributions.

Machine learning-based neural network potentials often cannot describe long-range interactions. Here the authors present an approach for building neural network potentials that can describe the electronic and nuclear response of molecular systems to long-range electrostatics.

The discovery of heterogeneous catalysts for large molecule conversion has been lagging due to the combinatorial inventory of intermediates. Here, the author presents an automated framework to explore the chemical space of reaction intermediates.

Theoretical description of light-matter coupling in the strong-coupling regime is challenging. Here the authors introduce a fully consistent ab-initio method of molecular orbital theory applicable to material systems in quantum electrodynamics environments.

Exploration of metastable phases of a given elemental composition is a data-intensive task. Here the authors integrate first-principles atomistic simulations with machine learning and high-performance computing to allow a rapid exploration of the metastable phases of carbon.