Machine Learning Force Fields

• Accurate force calculations are impractical for large systems: CCSD(T) scales ∝N^7, so molecular dynamics commonly uses efficient but less accurate empirical force fields.Classical force fields derive atomic forces analytically and enable simulations of many thousands of atoms.Source: §1 · pp. 2–3
• Conventional force fields inadequately model polarization and many-body interactions, require predefined bonding patterns, and often cannot quantitatively describe bond breaking or formation.Reactive force fields exist, but they are often insufficiently accurate for quantitative studies or limited to particular reaction types.Source: §1 · p. 3
• Machine-learning methods aim to close the gap between ab initio accuracy and classical-force-field efficiency by learning structure–property relationships from data rather than solving governing equations directly.This approach can reflect effective quantum-mechanical rules while exploring chemical space with high efficiency and accuracy.Source: §1 · p. 4
• ML force fields learn energies and forces from ab initio reference data without preconceived bonding patterns, reconstructing interactions from atomic positions and naturally describing chemical reactions.Small-molecule ML force fields trained on CCSD(T) data can approach spectroscopic accuracy with computational efficiency similar to conventional force fields.Source: §1 · p. 4
• The review covers ML-force-field fundamentals, construction and testing best practices, applications yielding new chemical insights, and remaining obstacles to broader applicability.It also discusses software packages and applications including molecular dynamics and related prediction targets or simulation enhancements.Source: §1 · pp. 4, 6, Figure 2 · p. 5

2.1 Chemical concepts

• It reviews potential energy surfaces and physical invariance properties as essential principles for constructing physically meaningful ML force fields.The discussion targets ML-focused readers seeking fundamental chemical knowledge.Source: §2 · p. 6

2.2 Machine-learning methodologies

• It introduces kernel-based learning and artificial neural networks as two important machine-learning methodologies for ML force fields.This overview provides chemical-background readers with an initial orientation to the relevant methods.Source: §2 · p. 6

2.3 Invariance constraints and models

• It presents constraints associated with physical invariances and examples of ML force-field models that implement those constraints in practice.These examples connect the conceptual invariance requirements to concrete model-construction approaches.Source: §2 · p. 6

2.1.1 The potential energy surface

• The potential energy surface maps nuclear charges and positions to energy, with minima representing stable structures and transition states connecting configurations through chemical reactions.Its topology governs molecular evolution, reaction pathways, product formation, and thermodynamic behavior.Source: §2.1.1 · pp. 6–7, Figure 3 · p. 7
• Forces are the negative position gradients of potential energy and enable molecular-dynamics integration, geometry optimization, and reaction-barrier calculations.Classical nuclear dynamics can be insufficient when nuclear delocalization is significant; nuclear quantum effects and path-integral molecular dynamics may then be required.Source: §2.1.1 · p. 7
• ML force fields learn analytical representations of the potential energy surface from reference data because ab initio energies and forces are often too costly for every molecular-dynamics step.This shifts the challenge from solving the electronic-structure equation to selecting and learning an appropriate functional form.Source: §2.1.1 · p. 7

2.1.2 Invariances of physical systems

• Physically meaningful ML force fields must preserve energy, roto-translational, and permutation invariance through constraints matching those of the true potential energy surface.Energy conservation requires forces to be negative energy gradients, while linear and angular momentum conservation implies dependence on relative atomic positions rather than rigid translations or rotations.Source: §2.1.2 · p. 8

2.2 Machine learning foundations

• ML force fields aim to generalize to unseen data when finite training sets cannot capture all information needed for exact reconstruction.Although ab initio reference data and smooth potential-energy surfaces can satisfy the assumptions of noise-free interpolation, insufficient data makes generalization the practical objective; Occam’s razor, bias–variance control, and cross-validation help select models that generalize.Source: §2.2 · p. 8
• Nonlinear energy prediction is often linearized by mapping inputs into higher-dimensional feature spaces, which kernels access implicitly while deep networks learn through layered nonlinear transformations.Kernel methods use the kernel trick, whereas neural networks compose linear transformations with nonlinear activations to learn complex feature spaces.Source: §2.2 · pp. 8–9, Figure 4 · p. 9

2.2.1 Kernel-based methods

• Kernel regression represents predictions as weighted sums of similarity functions over structural descriptors such as internal coordinates, Coulomb matrices, atomic environments, or crystal encodings.The kernel measures similarity between descriptors, and Gaussian kernels can operate in an implicitly infinite-dimensional feature space through the kernel trick.Source: §2.2.1 · pp. 9, 11, 12
• Regularization smooths kernel predictions, improves numerical stability and generalization, and trades exact reproduction of reference values for reduced overfitting.For kernel regression, adding λ > 0 damps coefficient magnitudes and can improve predictions on unknown values; the choice of kernel also imposes assumptions about function smoothness and restricts the hypothesis space.Source: Figure 6 · p. 10, §2.2.1 · p. 12
• Kernel methods can become computationally expensive at scale, with exact solution costs of O(M^3), while iterative and low-rank approximations reduce training or evaluation complexity at the cost of approximation or reduced capacity.Iterative solvers reduce complexity to O(M^2), subset methods scale as O(ML^2) for solving and O(L) for evaluation, and Nyström approximations scale as O(L^3 + ML^2).Source: §2.2.1 · pp. 12–13

2.2.2 Artificial neural networks

• Neural networks require nonlinear activations across stacked layers to model nonlinear mappings, and deep architectures are often more parameter-efficient than shallow ones.A sufficiently wide hidden layer gives general function approximation, while multiple hidden layers can achieve comparable expressiveness with fewer parameters.Source: §2.2.2 · p. 15
• For potential-energy surfaces, neural networks map chemical-structure representations to scalar energies, learning weights and biases by stochastic optimization of a loss such as mean squared error.Unlike kernel coefficients, neural-network parameters cannot be fitted in closed form and are optimized from randomly sampled training batches.Source: §2.2.2 · p. 15
• Neural-network evaluation cost is independent of training-set size, unlike kernel-method fitting and evaluation costs, making neural networks suitable for large datasets.Kernel fitting scales as O(M^3) and evaluation as O(M), whereas neural-network evaluation scales with the chosen architecture.Source: §2.2.2 · p. 15

2.2.3 Model selection: How to choose hyperparameters

• Model hyperparameters govern architecture or training and are selected using methods such as grid, random, or Bayesian search because their domains and dependencies vary.Examples include neural-network width and depth, kernel width γ, and regularization strength λ; performance is often robust to small changes.Source: §2.2.3 · pp. 15–16
• An independent test set must remain untouched during hyperparameter selection, while training-validation splits or k-fold cross-validation estimate generalization before final test evaluation.Validation performance can be optimistic because it influences model selection; test performance is reliable only when no further tuning responds to it.Source: §2.2.3 · p. 16

2.3 Combining machine learning and chemistry

• ML methods in chemistry shortcut computationally complex physical calculations through empirical inference, while force-field construction must remain accurate despite limited high-level training data.The central challenge is retaining generality, generalization, versatility, transferability, and scalability with datasets much smaller than those common in traditional machine-learning domains.Source: §2.3 · p. 17

2.3.1 Physical constraints

• ∼1.5 times the energy-only cost is sufficient for PBE0/DFT analytic energies and forces, versus at least ∼10 times for numerical gradients, while direct-force models require explicit energy-conservation constraints.Energy-gradient models are conservative by construction; direct-force models exploit analytically available force observables and avoid derivative-error amplification but must represent a valid gradient field.Source: §2.3.1 · pp. 17–18, Figure 10 · p. 18

Indistinguishability of identical atoms

• Roto-translational and permutation invariance are essential because violating them can produce spurious dynamics or different predictions for physically equivalent arrangements of identical atoms.Weak assignment schemes may introduce discontinuities during molecular dynamics, whereas graph matching is costly and can degrade accuracy near symmetry-domain boundaries.Source: §2.3.1 · p. 18, Indistinguishability of identical atoms · p. 19

2.3.2 (Symmetric) Gradient Domain Machine Learning ((s)GDML)

• GDML predicts forces directly with a differentiated energy kernel, and sGDML adds data-driven rigid and non-rigid symmetry constraints to obtain conservative, permutation-invariant models efficiently.Force training provides 3NM values from M reference structures, while symmetry matching excludes physically irrelevant permutations and reduces learning complexity without additional approximations.Source: §2.3.2 · pp. 20–21

2.3.3 Gaussian approximation potentials (GAPs)

• GAPs decompose molecular energies into local atom-centered contributions, learn that decomposition from whole-system reference energies, and use descriptors such as SOAP that preserve rotational and neighbor-permutation invariance.GAPs scale linearly with atom number and can accommodate periodic boundary conditions.Source: §2.3.3 · p. 21

2.3.4 Neural Network Potentials

• Neural-network potentials address variable system sizes by summing shared atom-wise energy contributions, while descriptor-based and end-to-end models encode environments through fixed features or learned message passing.Transferability requires sufficient local-environment sampling and long-range corrections; end-to-end architectures include MPNNs such as DTNN, SchNet, HIP-NN, and PhysNet.Source: §2.3.4 · pp. 22–23
• Energy-based NNPs obtain conservative forces by differentiation and can incorporate ab initio forces during training, but smooth activation functions are required to avoid force discontinuities.The model representations must remain invariant to translation, rotation, and permutation of symmetry-equivalent atoms; popular nonsmooth ReLU activations are therefore excluded.Source: §2.3.4 · p. 23
• Descriptor-based NNPs encode local environments with manually designed symmetry functions, whose element-wise duplication can become combinatorially expensive in multicomponent systems.These descriptors use internal geometric information, cutoff functions, and parameterized distance- and angle-sensitive terms to predict atomic energy contributions.Source: §2.3.4 · pp. 23–24
• End-to-end NNPs learn environment representations from reference data, avoiding descriptor-choice limitations while retaining roto-translational and permutational invariance through distance-based message passing.Learned features adapt to training-set size, distribution, and chemical property, while commutative aggregation and pairwise distances enforce the relevant invariances.Source: §2.3.4 · p. 24
• Including explicit electrostatic and dispersion terms improved energy and force accuracy for SN2 reactions, whereas omitting them produced qualitatively wrong asymptotic behavior.PhysNet predicts corrected partial charges constrained to sum to the molecule’s total charge and adds Grimme’s D3 dispersion correction.Source: §2.3.4 · p. 25

3 Best practices and Pitfalls

• The workflow covers preliminary considerations, ML-method selection, reference-data collection and preparation, model training, production use, and software packages with code examples.These steps are organized across Sections 3.1–3.7 and include attention to limitations of ab initio methods and problems arising when recommended practices are not followed.Source: §3 · p. 25

Preliminary considerations

• The limitations of the chosen level of theory are problem-specific and affect potential-energy-surface reconstruction generally, not only ML-FFs.A comprehensive assessment is therefore impossible, but these examples illustrate key preliminary checks.Source: §3.1 · p. 25

Practicability

• ML-FF construction is worthwhile only when total data-generation, training, testing, and inference costs provide time savings over ab initio simulation and improved accuracy over conventional force fields.ML-FFs are one to three orders of magnitude slower than classical force fields and remain bounded by the accuracy of their reference data.Source: §3.1 · pp. 25–26

Multireference effects

• Multireference effects can cause discontinuous electronic-state switches across distorted geometries, producing inconsistent reference data that degrade ML-model performance.Such effects should be checked before data generation, with multireference methods used when necessary.Source: §3.1 · p. 26

Strong delocalization

• Strong delocalization violates the locality assumed by many ML-FF models, particularly for extensive conjugated π-systems with non-local electronic interactions.Mitigation may require a larger cutoff radius or a different structural descriptor.Source: §3.1 · p. 26

Reference-data availability

• Prefer kernel methods below ∼10^3–10^4 training points, but use NNs for larger datasets because kernel training and evaluation become costly with dataset size.Kernel models often perform well with few data and permit analytical parameter optimization, whereas NN evaluation can remain constant-complexity; kernel training scales O(M^3) with O(M^2) memory and evaluation scales O(M).Source: §3.2 · p. 27

System variability

• For variable-size or multi-system targets, use models with atomic-contribution decompositions; fixed-length descriptors such as those in (s)GDML restrict applicability to a single system type.Most NNPs and many kernel methods, including GAPs and FCHL, support differently sized systems, while inverse-distance descriptors change length with atom number.Source: §3.2 · p. 27

Elemental diversity

• When many elements are covered, element embeddings are more scalable than descriptors whose terms grow combinatorially with the number of atom types.Element embeddings may also improve data efficiency by exploiting alchemical information, whereas separate element-specific geometric terms increase descriptor size, computational cost, and data requirements.Source: §3.2 · p. 28

System size

• For very large systems, local cutoff models enable fragment-based training and parallel atomic predictions, although message-passing models require communication across iterations or larger replicated subdomains.Locality can make calculations feasible when full-system ab initio data are impractical and when all atoms cannot fit in one machine’s memory.Source: §3.2 · pp. 28–29

Interaction range

• If strong long-range interactions matter, avoid cutoffs or add explicit physical interaction terms, because cutoff models can incur significant errors beyond the cutoff.Increasing the cutoff reduces computational and data efficiency; TensorMol and PhysNet include correction terms by default, and such terms may improve data efficiency even without cutoffs.Source: §3.2 · p. 29

Data collection

• Reference-data quality and coverage define an ML-FF’s reliability and applicability, but expensive calculations and vast configurational spaces make representative sampling difficult.Deficiencies in the data propagate into model artifacts, so sampling choices may need to match the model and intended application.Source: §3.3 · p. 29

AIMD sampling

• AIMD offers straightforward PES exploration, with temperature controlling sampled regions and energies, but its low-energy bias suits equilibrium or near-equilibrium properties better than rare events.Examples include vibrational spectra, minima populations, and thermodynamic properties.Source: §3.3 · pp. 29–30, Figure 14 · p. 30

Sampling by proxy

• Sampling by proxy reduces expensive high-level calculations by subsampling a long low-level trajectory, but it is reliable only when the two PESs have similar topology.Topologically different sampling levels can miss relevant high-level regions and cause unphysical ML-FF predictions during MD.Source: §3.3 · pp. 30–31, Figure 15 · p. 30

Adaptive and metadynamics sampling

• Adaptive sampling augments training data when uncertainty detects unreliable MD configurations, whereas metadynamics biases dynamics toward unexplored PES regions.Uncertainty estimates must balance missed deviations against redundant calculations; query-by-committee provides an alternative for models lacking explicit uncertainty estimates.Source: §3.3 · pp. 31–32

Normal mode sampling

• Normal-mode sampling generates uncorrelated Boltzmann-like distortions efficiently, but its harmonic approximation restricts coverage near minima and worsens at higher temperatures.It is therefore best combined with other methods, such as adaptive sampling, to expand an initial dataset.Source: §3.3 · p. 32

Problems due to insufficient sampling

• 500 K sampling is recommended for a 300 K application because broader PES coverage provides a buffer against unreliable extrapolation and unphysical predictions.Models agree near sampled configurations but can diverge and behave unphysically outside them; extrapolation may also impair reaction dynamics and produce spurious minima or incorrect bond behavior.Source: §3.3 · p. 33

Splitting the data

• Keep training/validation and test sets strictly disjoint because leakage or correlated snapshots can make generalization-error estimates overly optimistic; time-splits may be preferable for long MD trajectories.Validation data supports hyperparameter tuning, whereas the test set is reserved for evaluating unseen-data performance; for correlated trajectories, one example uses the first 80% for training/validation and the last 20% for testing.Source: §3.4 · p. 34

Data preprocessing

• Use only the training set to compute normalization or other data-dependent transformations, preventing test-set information from producing overconfident generalization estimates.Inputs or targets may be scaled and shifted to −1 … 1 or to zero mean and unit variance; mean-energy subtraction should likewise use training structures only.Source: §3.4 · p. 34
• When training on energies and forces jointly, scale corresponding labels consistently because forces derive from energy, whereas adding a constant to force labels breaks energy conservation.Multiplying energy labels by a factor requires applying the same factor to forces; subtracting a constant from energies is valid, but adding one to forces introduces an inconsistent linear energy term.Source: §3.4 · p. 34
• Unit conversions must preserve consistency across energies, coordinates, and forces, such as converting forces to kcal mol−1 Å−1 when energies and coordinates are converted to kcal mol−1 and Å.Raw reference data may mix atomic units for energies and forces with angstroms for coordinates, depending on the source code.Source: §3.4 · p. 34

3.5 Training the model

• Training tunes model parameters to minimize prediction loss, using analytical solutions when available or iterative optimization such as gradient descent for neural networks.Adam is a useful default optimizer because it converges quickly and performs well across many neural-network architectures.Source: §3.5 · p. 35

Choosing the loss function

• For regression, MSE is standard but disproportionately penalizes outliers, whereas MAE, Huber, or adaptive losses may be preferable in other cases.MSE assumes normally distributed reference noise and permits analytical optimization for convex linear models such as kernel ridge regression.Source: §3.5 · p. 35
• Hybrid energy–force losses weight both objectives through η, but because their optima generally differ, training may require a compromise rather than improve both predictions simultaneously.Force-only models can recover energy by integration, while model quality should ultimately be judged by experimental observables and generalization to unseen data.Source: §3.5 · pp. 35–36

Tuning hyperparameters

• Hyperparameters are selected by repeatedly training configurations and evaluating them on validation data, never test data, with grid or random search available for automation.Using the test set during tuning would compromise estimation of the model’s generalization error.Source: §3.5 · p. 36

Regularization

• Regularization limits model complexity because overfitting can produce low training error but poor unseen-data performance and spurious minima on the predicted PES.Early stopping uses the validation-loss minimum, while L2 and L1 penalties shrink parameters or favor sparse combinations; regularization strength λ is validation-tuned.Source: §3.5 · pp. 36–37

3.6 Using ML-FFs in production

• Production use should begin with test-set accuracy checks such as MAE or RMSE, which summarize overall performance and enable simple benchmarking.These measures estimate generalization because the test set was not used directly or indirectly during training.Source: §3.6 · p. 37
• Test-set summary errors can be misleading when the set is not representative of production data, because they emphasize densely sampled regions and may conceal extrapolation failures.For unfamiliar tasks or distributions, collecting additional reference points is recommended to verify continued validity.Source: §3.6 · p. 37
• Production trajectories should be inspected for physically sensible structures, including plausible bond lengths and valencies, because rare unsampled configurations can produce questionable predictions.Any suspicious prediction should be checked with additional reference calculations.Source: §3.6 · p. 38

3.7 Example code and software packages

• Reference implementations make intricate ML-FF details accessible, support replication of numerical experiments, and enable building on existing models with minimal effort.Source: §3.7 · p. 38

3.7.1 The sGDML package

• The sGDML package provides command-line training and testing from energy-and-force-labeled geometries, automatically performing model selection and hyperparameter tuning while supporting MD and other ASE/i-PI workflows.Force labels enforce energy conservation through an explicit linear operator constraint, and interfaces support vibrational analyses, optimizations, nudged elastic band computations, and simulations.Source: §3.7.1 · p. 38

3.7.2 The SchNetPack package

• SchNetPack244 supports extensible atomistic neural networks for energies, forces, and other properties, with ASE integration, benchmark datasets, GPU-enabled training, and molecular or path-integral MD.Its command-line tools handle custom or preexisting datasets, configurable training, model evaluation, and simulations with temperature, pressure, and sampling controls.Source: §3.7.2 · pp. 39–41

3.7.3 Other software packages

• Other packages cover diverse ML-FF needs, including ASE-integrated on-demand fitting in AMP, neural-network potentials in ænet, DeePMD-kit, n2p2, PROPhet, and TensorMol, and kernel workflows in Dscribe and QML.These tools vary in descriptors, frameworks, simulation interfaces, and supported physical interactions, including electrostatics in TensorMol.Source: §3.7.3 · pp. 41–42
• The RKHS toolkit reduces kernel-PES evaluation from linear scaling with training points M to O(log N) using lookup tables, but requires grid-structured data and is mainly suited to small systems.Its kernels can encode long-range decay and enable extrapolation beyond the training range.Source: §3.7.3 · p. 42

4.1 Electronic Effects

• ML-FFs capture electronic effects such as lone-pair interactions, hybridization changes, conjugation, and n →π∗ interactions that conventional force fields neglect or misrepresent, altering molecular dynamics and properties.These effects determine configurations, free-energy surfaces, infrared spectra, and reaction-relevant molecular geometries across ethanol, malondialdehyde, paracetamol, and aspirin.Source: §4.1 · pp. 43, 45

4.2 Thermodynamics

• ML-FF simulations reproduce water’s measured structural and transport properties, attribute its thermodynamic anomalies to weak van der Waals forces, and enable reactive surface and bulk-material studies.They reveal short-lived hydroxyl groups on titanium oxide, investigate defects and crack propagation, and predict phase diagrams requiring bond formation and breaking.Source: §4.2 · pp. 45–46

4.3 Reactions

• Up to three orders of magnitude lower errors and qualitatively different Helmholtz free energy surfaces were achieved by ML-FFs for the Cl–CH3–Br transformation versus classical methods.For methane activation by MgO+, ML-FF errors were up to two orders of magnitude lower, and predicted rate-constant disagreement with experiment was halved.Source: §4.3 · p. 46
• ML-FFs reduced the computational cost of acetaldehyde photoreaction trajectories by more than five orders of magnitude, showing that atmospheric ethenol formation is unlikely.The same model described tautomerization and both dissociation pathways without the prohibitive cost of approximately 1 µs of AIMD.Source: §4.3 · p. 46

4.4 Nuclear Quantum Effects

• Speed-ups of 10^5–10^7 enable ML-FF path-integral simulations that incorporate nuclear quantum effects, which can substantially change free energies, conformer populations, interactions, and proton-transfer rates.Experimentally determined ethanol conformer populations were recovered only with NQEs, while aqueous NaOH simulations showed increased proton-transfer rates and diffusion coefficients.Source: §4.4 · pp. 47, 49

4.5 Excited States

• Phase-less loss functions and proxy-potential models for non-adiabatic couplings enable accurate excited-state population statistics and longer simulations without costly preprocessing.Descriptor-based NNPs also recovered effective potential-energy surfaces near methylene-imine conical intersections with efficient surface-hopping simulations.Source: §4.5 · pp. 49–50

4.6 Spectroscopy

• Time autocorrelation functions from ML-FF molecular dynamics provide diffusion coefficients and infrared or Raman spectra, which connect molecular structure and dynamics to experiment.ML models extending beyond potential energies to dipoles and polarizabilities make such simulations substantially cheaper than ab initio calculations and can accommodate quantum effects such as vibrational anharmonicity.Source: §4.6 · p. 50
• 9000 years of projected ab initio computation was reduced to a few days for infrared spectra of systems including a protonated alanine tripeptide using ML-FFs (Ref. 71).The approach combined a descriptor-based NNP potential-energy model with latent neural-network-predicted atomic charges, while its accuracy exposed limitations in the reference methods.Source: §4.6 · p. 50
• Raman-spectrum ML models achieved excellent agreement with electronic-structure methods, used fewer than 1000 references, transferred across crystal polymorphs, and enabled 416-water-molecule, two-nanosecond simulations (Refs. 278, 279).Atomic-resolution decomposition separated intra- and intermolecular contributions, revealing mechanisms behind temperature-dependent spectral features.Source: §4.6 · p. 51

5 Challenges

• Larger ML-FF applications are limited because system complexity, conformational diversity, and ab initio costs grow rapidly with system size, restricting reference-data collection.The number of stable structures increases exponentially with atomistic degrees of freedom, while more atom correlations must be represented to capture full interactions.Source: §5 · p. 51

5.1 Locality and smoothness assumptions

• Locality and smoothness improve ML-FF efficiency and stability but conflict with many-body physics and fail for nonlocal, nonsmooth potential-energy surfaces such as cumulenes.Cumulene predictions increasingly deteriorate with chain length, and all models fail to reproduce the angular energy dependence for n = 7.Source: §5.1 · pp. 52–53, Figure 20 · p. 54

5.2 Transferability, scalability and long-range interactions

• Local architectures enable transferability, fragment-based training, and linear scaling, but struggle with conjugation, excited states, electrostatics, dispersion, and other long-range interactions.Increasing cutoffs can be data-inefficient, while global models require whole-system reference calculations and are less reusable across systems.Source: §5.2 · pp. 53, 55, 56
• Promising remedies combine ML with physical long-range treatments, electronic-structure representations, or hybrid ML/MM embeddings, trading broader interaction recovery against computational cost or embedding limitations.Examples include ML-predicted charges with Coulomb or Ewald summation, semi-empirical ML, Hamiltonian models, and explicitly polarized ML regions in classical environments.Source: §5.2 · pp. 56–57

Review contributions

• The review covers physical invariances, ML algorithms, loss functions, careful validation, best practices, pitfalls, challenges, and software tools for constructing and adopting ML force fields.The guidance explicitly spells out often-omitted practical know-how to reduce barriers for researchers entering the field.Source: §6 · pp. 57–58

Achievements

• A few hundred ab initio calculations can produce ML force fields for aspirin dynamics at coupled-cluster accuracy within a few tens of wavenumbers, while ML speed-ups enable laptop-scale studies.These methods also automate force-field construction, replacing labor-intensive human effort.Source: §6 · p. 58

Remaining limitations

• Condensed-phase simulations at the highest levels of theory remain beyond reach, despite ML force fields already making previously impractical molecular-dynamics studies possible.The remaining limitation qualifies the field’s current achievements and scope.Source: §6 · p. 58

Chemical insights

• ML force fields have revealed non-trivial electronic effects in relatively small molecules and may enable further chemical discoveries as larger-system studies become feasible.The review connects these new studies to improved understanding of experimental observations.Source: §6 · p. 58

Outlook

• ML force fields remain an active interdisciplinary research area with unexplored applications and opportunities for quantum-chemical, algorithmic, and practical advances.The review expects ML force fields to become a fundamental part of modern computational chemistry.Source: §6 · p. 58

Authors

• Oliver T. Unke develops accurate potential energy surfaces and applies them to molecular dynamics simulations.He is an SNSF postdoctoral research fellow at Technische Universität Berlin and received his Ph.D. in Chemistry from the University of Basel in 2019.Source: Biographies · p. 58
• The authors’ research spans data-efficient and robust quantum-chemistry learning, ab initio simulations, nuclear quantum effects, thermodynamics, and machine-learning applications in materials science.These interests are represented by Stefan Chmiela, Huziel E. Sauceda, Michael Gastegger, and Igor Poltavsky.Source: Biographies · pp. 58–59
• Kristof T. Schütt studies interpretable neural networks, representation learning, generative models, and machine-learning applications in quantum chemistry.He is a senior researcher at BIFOLD and previously developed neural networks for real-time speech enhancement at Audatic.Source: Biographies · p. 59
• Alexandre Tkatchenko leads research connecting quantum mechanics, statistical mechanics, and machine learning to accurate modeling of complex materials.He is Professor of Theoretical Chemical Physics at the University of Luxembourg and Visiting Professor at Technische Universität Berlin.Source: Biographies · p. 59
• Klaus-Robert Müller directs major Berlin machine-learning centers and researches intelligent data analysis and machine learning in neuroscience, physics, chemistry, and industry.He has been a professor of computer science at Technische Universität Berlin since 2006.Source: Biographies · pp. 59–60