PaperlayerView exact revision on arXiv

Source-linked AI summary

Perspective on integrating machine learning into computational chemistry and materials science

Julia Westermayr, Michael Gastegger, Kristof T. Schütt, Reinhard J. Maurer

arXiv:2102.08435v3physics.chem-ph

TL;DR

Computational chemistry and materials science face a rapidly expanding landscape of machine-learning models whose broader practical integration remains incomplete. This perspective surveys ML across electronic-structure and simulation workflows, highlighting progress toward faster, more transferable modeling while identifying barriers to automation, efficiency, and openness.

  • Problem

    Few ML models are yet generally applicable beyond their developer communities, making it difficult for computational molecular scientists to identify dependable approaches and integrate them into standard workflows.

  • Method

    The paper provides an accessible perspective on recent ML advances across model building, electronic structure, molecular dynamics, and spectroscopy, emphasizing their integration into workflows and software.

  • Results

    ML approaches now support faster simulations, data-efficient electronic-structure modeling, active sampling for molecular dynamics, and direct integration with quantum-chemistry calculations.

  • Takeaways & Limitations

    The perspective identifies permanent integration of ML into computational workflows and software as a central opportunity for improving the scale, transferability, and usability of atomistic modeling.

  • Takeaways & Limitations

    Complete automation of ML model building has not yet been achieved, and computational efficiency remains a challenge as simulations target larger systems and longer timescales.

Abstract

from arXiv · show

Machine learning (ML) methods are being used in almost every conceivable area of electronic structure theory and molecular simulation. In particular, ML has become firmly established in the construction of high-dimensional interatomic potentials. Not a day goes by without another proof of principle being published on how ML methods can represent and predict quantum mechanical properties - be they observable, such as molecular polarizabilities, or not, such as atomic charges. As ML is becoming pervasive in electronic structure theory and molecular simulation, we provide an overview of how atomistic computational modeling is being transformed by the incorporation of ML approaches. From the perspective of the practitioner in the field, we assess how common workflows to predict structure, dynamics, and spectroscopy are affected by ML. Finally, we discuss how a tighter and lasting integration of ML methods with computational chemistry and materials science can be achieved and what it will mean for research practice, software development, and postgraduate training.

I. INTRODUCTION

Machine learning is rapidly entering electronic-structure and atomistic-simulation workflows, with potential to accelerate established methods and enable new workflows. This perspective examines that transformation from the practitioner’s viewpoint, emphasizing integration into software, research practice, and training.

  • Quantum-based atomistic simulations support first-principles property prediction, atomic-scale dynamics, mechanistic understanding, materials discovery, and materials optimization.
  • Recent ML work spans catalyst design, interatomic potentials, quantum chemistry, Schrödinger-equation solutions, and unsupervised learning for atomistic simulation.
  • ML models can parameterize analytical electronic-structure models, enabling very fast evaluation and potentially longer simulation time and length scales.
  • The perspective asks how ML will affect computational workflows, practitioner expertise, and the training requirements of future PhD graduates.
  • Its goal is lasting integration of ML and simulation through shared code and data structures or bidirectional data exchange, rather than a comprehensive review of existing approaches.

II. MACHINE LEARNING PRIMER

The primer introduces supervised, unsupervised, generative, and reinforcement-learning concepts relevant to molecular modeling. It emphasizes that representation, training-data coverage, and physical prior knowledge govern accuracy, data efficiency, and extrapolation reliability.

  • Supervised learning maps inputs to labelled targets for regression or classification, while unsupervised learning uses unlabeled data for tasks such as dimensionality reduction or clustering.
  • Training examples should represent the application distribution because extrapolation beyond the training domain quickly makes predictions unreliable.
  • Encoding physical or structural prior knowledge, such as invariances or baseline differences, reduces the effective input space and required training data.
  • Generative models can learn probability distributions over molecular structures conditioned on desired property ranges, supporting inverse design.
  • Reinforcement learning selects actions that maximize future rewards and can support molecular design without a representative reference-structure set before training.

III. ML IMPROVES MODEL BUILDING, METHOD CHOICE, AND OPENS NEW MULTI-SCALE APPROACHES

ML can assist model construction and method choice while enabling multi-scale approaches that combine different levels of theory. These developments improve access to balanced modeling decisions and can make otherwise infeasible simulations possible, but full automation remains unresolved.

  • III. ML IMPROVES MODEL BUILDING, METHOD CHOICE, AND OPENS NEW MULTI-SCALE APPROACHES: Model building balances chemical accuracy against computational feasibility through structural-model design and computational-method choice.
  • III. ML IMPROVES MODEL BUILDING, METHOD CHOICE, AND OPENS NEW MULTI-SCALE APPROACHES: Because modeling choices are ambiguous and depend on expert chemical intuition, ML can learn decision rules to support more widely available and potentially automated model selection.
  • III. ML IMPROVES MODEL BUILDING, METHOD CHOICE, AND OPENS NEW MULTI-SCALE APPROACHES: Uncertainty quantification can make method-selection protocols transparent by exposing confidence intervals rather than reducing predictions to single numbers.
  • III. ML IMPROVES MODEL BUILDING, METHOD CHOICE, AND OPENS NEW MULTI-SCALE APPROACHES: ML has been used to construct basis sets and local pseudopotentials, extending automation beyond choosing among existing electronic-structure methods.
  • III. ML IMPROVES MODEL BUILDING, METHOD CHOICE, AND OPENS NEW MULTI-SCALE APPROACHES: ML multi-scale models can reproduce higher-level accuracy, accelerate solvent calculations by up to four orders of magnitude, and enable crack-propagation simulations otherwise impossible with a single modeling level.
  • III. ML IMPROVES MODEL BUILDING, METHOD CHOICE, AND OPENS NEW MULTI-SCALE APPROACHES: Complete automation of model building has not yet been achieved, leaving automated correlation-method selection and multi-scale partition generation as future directions.

IV. ML IN ELECTRONIC STRUCTURE THEORY

ML addresses computational bottlenecks in electronic structure theory by replacing or augmenting expensive calculations with learned representations, while improving data efficiency and software integration. Its scope remains constrained by non-smooth excited-state targets, limited transferability, and the need for modular workflows.

  • Motivation: Electronic structure calculations face bottlenecks from evaluating multi-centre and multi-electron quantities and iteratively solving coupled equations.ML is positioned as especially impactful for the first challenge, while scalable linear algebra libraries have advanced the second.
  • Interatomic potentials: ML interatomic potentials can replace ab-initio calculations while retaining ab-initio accuracy at costs comparable to classical force fields.This is currently the most pervasive ML application in electronic structure and improves the predictive capabilities of molecular dynamics.
  • Limitations: ML models for excited-state properties face difficult learning targets because avoided crossings and conical intersections create non-smooth energies, cusps, and singular couplings.Higher-lying excited states are especially challenging when multi-reference methods are required because state-character switches produce jumps and noisy potential-energy surfaces.
  • Data-efficient representations: Physics-informed representations improve data efficiency and transferability, with MOB-ML reaching chemical accuracy using 36 times fewer data points than Δ-ML for molecules up to 13 heavy atoms.Related approaches include symmetry-aware neural networks, physically constrained kernels, and ML parametrizations of DFTB.
  • Integrated electronic-structure models: SchNOrb predicts Hamiltonians and overlap matrices in local atomic orbital representations compatible with quantum chemistry software, reducing self-consistent-field iterations by an average of 77%.Its outputs can serve as wave-function initial guesses or enter perturbation-theory calculations, while localized effective minimal bases benefit larger-system accuracy.
  • Software integration: Tighter ML integration depends on modular electronic-structure software, interoperable libraries, universal data standards, and scalable multi-language communication.The perspective identifies interfaces among ML, electronic-structure, dynamics, algebra, and data-repository tools as a route toward integrated ML/QM workflows.

V. ML WILL IMPROVE OUR ABILITY TO EXPLORE MOLECULAR STRUCTURE AND MATERIALS COMPOSITION

ML methods expand molecular and materials exploration from chemical composition and structure to global energy landscapes, local reaction pathways, and transition states. They improve optimization, transition-state searches, global structure discovery, and property-directed molecular generation.

  • Exploration spans chemical space, global searches on fixed-composition potential-energy surfaces, and local searches for reaction pathways and transition states.These levels vary chemical composition and structure, structural conformations and stability, or local potential-energy-surface details.
  • ML-based preconditioning reduces the number of geometry-optimization steps required for molecules, transition-metal complexes, correlated quantum chemistry, bulk materials, and adsorbed molecules.
  • A factor of 2 speedup was reported for one-ended transition-state searches using Gaussian-process regression compared with conventional methods.
  • 5 to 25 fewer energy and force evaluations were required by a surrogate Gaussian-process model accelerating nudged elastic band searches compared with conventional NEB.
  • ML combined with structure optimization has enabled major advances in difficult global-search problems, including protein folding and crystal, surface, and interface structure prediction.
  • Generative ML can propose chemically valid molecules with targeted properties, but graph-based models cannot distinguish conformations sharing the same molecular graph.
  • ML models can represent alchemical potentials and generate smooth paths through alchemical space, supporting future continuous variation of elemental composition to optimize material properties.

AND COMPLEXITY

ML extends atomistic simulation across classical, mixed quantum-classical, and quantum dynamics by replacing expensive evaluations, learning reduced representations, and identifying relevant dynamical coordinates. These gains improve accessible timescales and sampling, but ML simulations remain slower than empirical force fields and face efficiency bottlenecks.

  • ML-based interatomic potentials replace on-the-fly electronic-structure evaluations and are now commonly established for molecular dynamics simulations.
  • Active learning samples relevant configuration space efficiently and can detect holes in potential-energy surfaces during production runs.
  • Gradient-domain and Δ-learning models provide energy-conserving, data-efficient force fields and can transfer higher-level accuracy from lower-level reference data.
  • ML identifies collective variables associated with long-time dynamics, helping characterize attractor states and explore hierarchical energy landscapes.
  • ML models represent coarse-grained potentials and free-energy surfaces for oligomers, liquid water, alanine dipeptide, and molecular liquids.
  • ML supports nonadiabatic and quantum dynamics through excited-state landscapes, electronic-friction tensors, automated diabatic representations, and reduced-dimensional diabatic potential-energy surfaces.
  • 100 femtoseconds of MQCD dynamics for CH2NH+ took 24 seconds with ML potentials versus 74,224 seconds with MR-CISD/aug-cc-pVDZ, while Amber required 0.005 seconds for classical dynamics.
  • ML dynamics remain computationally challenging for long timescales and large ensembles, motivating sparsity and efficient high-body-order polynomial representations.

VII. ML HELPS TO CONNECT THEORY AND EXPERIMENT

ML connects theory and experiment by predicting experimentally comparable spectra, extracting molecular and materials information from measurements and literature, and guiding discovery and inverse design. However, exhaustive screening and reliable transfer across experimental conditions remain unresolved boundaries.

  • Computational spectroscopy uses ML to predict vibrational and response properties that can be directly compared with experiments.
  • A single field-dependent ML model predicted IR, Raman, and NMR spectra while modeling solvent effects through the molecular environment.
  • ML can infer structural and electronic information from spectroscopy, including functional groups and atomic structure from experimental measurements.
  • Natural-language processing extracted structure–property relationships from research literature, generalized learned concepts, and recommended materials for functional applications.
  • A model trained on failed crystallization experiments predicted reaction success, learned general reaction conditions, and revealed hypotheses for successful product formation.
  • ML combined with high-throughput screening supports molecular, drug, catalyst, perovskite, and polymer discovery.
  • Inverse design reverses property prediction by generating structures with desired properties, but chemical space exceeds 10^60 molecules, making exhaustive screening infeasible.
  • Open issues include reproducing varying experimental conditions such as solvents and electromagnetic fields, improving data efficiency, and ensuring reliable data availability.

VIII. OUTLOOK

The outlook calls for ML to become a sustained, integrated component of electronic-structure and molecular-simulation software, supported by usable infrastructure, open data, and long-term community effort. This integration could extend computational capabilities, reshape software design, and broaden participation in computational research.

  • Software integration: ML integration could make electronic-structure and molecular-simulation software more computationally efficient through faster integral evaluation, improved initial guesses, and descriptions of nonlocal effects.The authors identify these as examples of how ML may augment existing computational techniques.
  • Software integration: Favorable ML scaling could extend molecular quantum dynamics simulations beyond currently infeasible timescales and system sizes.The passage identifies the present boundary as a few picoseconds and tens of atoms.
  • Implementation: User-friendly, modular, well-maintained software is necessary for tight integration with established deep-learning platforms and chemistry codes.The authors emphasize modularity and interfaces to platforms such as TensorFlow and PyTorch.
  • Community infrastructure: Training-data availability and willingness to share data and ML models remain crucial challenges, requiring standards and repositories while balancing commercial interests.The perspective cites FAIR-DI, NOMAD, Materials Project, and MolSSI QCArchive as examples of needed infrastructure.
  • Community infrastructure: Sustainable integration will require long-term community effort and continued funding rather than relying only on proof-of-principle applications.The authors specifically call on funding agencies, reviewers, and industrial stakeholders to support sustained efforts.
  • Research practice: Integrated ML could prompt reconsideration of established software design choices, including basis representations historically selected for computational efficiency.The authors use Gaussian basis functions and multicentre integrals as an example.
Loading 2102.08435v3…