Source-linked AI summary
Machine Learning in QM/MM Molecular Dynamics Simulations of Condensed-Phase Systems
Lennard Böselt, Moritz Thürlemann, Sereina Riniker
TL;DR
QM/MM simulations are costly, while machine-learned models face challenges in condensed-phase systems with long-range interactions and demanding sampling. The paper develops an electrostatically embedded HDNNP/MM workflow and extends it to delta-learning; its final models produce stable trajectories with accuracy close to the DFT reference in two molecular systems.
Problem
QM/MM is computationally costly, and condensed-phase ML models face challenges from long-range interactions, flexible-system sampling, and training demands.
Source: §1 · pp. 2, 4Method
The workflow incorporates MM particles into an HDNNP for electrostatic embedding and extends the model to learn the difference between DFTB and a reference method.
Source: §1 · pp. 4–5Results
Stable trajectories for retinoic acid in water and the SAM–cytosine transition state were obtained with accuracy close to the DFT reference.
Source: §5 · p. 27Takeaways & Limitations
The reported results indicate that delta-learning is a promising approach for (QM)ML/MM MD simulations of condensed-phase systems.
Source: §5 · p. 27Delta-learning is limited for configurations where the DFTB potential-energy surface differs substantially from the reference DFT surface.
Source: §4.3 · p. 26
Abstract
from arXiv · showhide
Quantum mechanics/molecular mechanics (QM/MM) molecular dynamics (MD) simulations have been developed to simulate molecular systems, where an explicit description of changes in the electronic structure is necessary. However, QM/MM MD simulations are computationally expensive compared to fully classical simulations as all valence electrons are treated explicitly and a self-consistent field (SCF) procedure is required. Recently, approaches have been proposed to replace the QM description with machine learned (ML) models. However, condensed-phase systems pose a challenge for these approaches due to long-range interactions. Here, we establish a workflow, which incorporates the MM environment as an element type in a high-dimensional neural network potential (HDNNP). The fitted HDNNP describes the potential-energy surface of the QM particles with an electrostatic embedding scheme. Thus, the MM particles feel a force from the polarized QM particles. To achieve chemical accuracy, we find that even simple systems require models with a strong gradient regularization, a large number of data points, and a substantial number of parameters. To address this issue, we extend our approach to a delta-learning scheme, where the ML model learns the difference between a reference method (DFT) and a cheaper semi-empirical method (DFTB). We show that such a scheme reaches the accuracy of the DFT reference method, while requiring significantly less parameters. Furthermore, the delta-learning scheme is capable of correctly incorporating long-range interactions within a cutoff of 1.4 nm. It is validated by performing MD simulations of retinoic acid in water and the interaction between S-adenoslymethioniat with cytosine in water. The presented results indicate that delta-learning is a promising approach for (QM)ML/MM MD simulations of condensed-phase systems.
1 Introduction
Machine-learned potentials could reduce the cost of QM/MM simulations, but condensed-phase applications face challenges from long-range interactions, demanding sampling, and model complexity. This work proposes an electrostatically embedded HDNNP/MM workflow and explores delta-learning for condensed-phase simulations.
1 Introduction
- QM/MM provides an alternative when electronic-structure changes matter or reliable force-field parameters are unavailable, but explicit electrons and SCF make the QM zone a computational bottleneck.Classical force fields neglect electron rearrangement in averaged properties.Source: §1 · pp. 1–2
- HDNNPs face steep sampling demands for flexible systems, struggle with long-range electrostatics, and can require costly training sets and many network weights.HDNNP descriptors also scale exponentially with the number of element types.Source: §1 · p. 4
- The proposed hybrid uses an HDNNP for QM particles and incorporates the MM environment into its descriptors, enabling electrostatic embedding and larger cutoffs.The MM environment is included in the QM/MM reference calculations used to generate training data.Source: §1 · p. 4
- The paper extends this workflow to delta-learning, training an ML model on the difference between a semi-empirical method and a higher-level reference.The introduction describes prior delta-learning work and states that the approach is applied here with a semi-empirical method.Source: §1 · pp. 3, 5
- The study explores model parameters and validates the approach in MD simulations of retinoic acid in water and SAM–cytosine in water, comparing results with standard QM/MM.The systems contain 50 and 63 QM-zone atoms, respectively.Source: §1 · p. 5
2 Theory
The theory contrasts quantum, semiempirical, and classical descriptions, then develops QM/MM and machine-learning approaches for representing molecular potential-energy surfaces. It culminates in an HDNNP–force-field framework that represents the QM region while retaining MM-environment effects.
2.1.1 Choice of the Model
- DFT and ab initio methods treat electronic structure explicitly, while DFTB and classical force fields offer approximate alternatives with different accuracy and computational demands.DFT requires all valence electrons and an SCF procedure; DFTB uses a truncated density expansion with fitted corrections, while force fields use effective parameters.Source: §2.1.1 · pp. 5–6
2.1.2 Classical Force Fields
- Classical force fields sum bonded and nonbonded terms, are inexpensive, and can reproduce bulk properties, but averaged electronic effects prevent their use for chemical reactions.Their parameters may incorporate long-range effects implicitly through fitting, and methods can include interactions beyond a given cutoff.Source: §2.1.2 · p. 6
2.1.3 QM/MM Scheme
- QM/MM combines a quantum subsystem with a classical environment; electrostatic embedding lets MM charges perturb QM particles, which in turn exert forces on the MM region.Converging accurate results can require a charge cutoff of 1.4 nm, while truncation makes the potential-energy surface discontinuous at the cutoff.Source: §2.1.3 · pp. 6–7
2.2 Machine Learning of Potential-Energy Surfaces
- Supervised ML fits a parameterized mapping from transformed configurations to estimated energies, optionally targeting a difference between theory levels or including gradient information.The input transformation must preserve information relevant to the target output.Source: §2.2 · pp. 7–8
2.2.1 High-Dimensional Neural Network Potentials (HDNNP)
- HDNNPs use element-specific networks and symmetry-function descriptors to represent atomic environments while enforcing invariance to atom swaps, rotations, and translations.Their limitations include training-data demands, limited extrapolation, and symmetry-function angle terms that scale exponentially with the number of element types.Source: §2.2.1 · pp. 8–9
2.3 (QM)ML/MM MD Simulations Using HDNNP
- The proposed QM/MM HDNNP learns the QM contribution while the classical force field supplies other terms; MM particles enter the descriptor as a charge-weighted element type.The MM environment is included in the training data, and point charges incorporate long-range interactions while allowing larger training cutoffs.Source: §2.3 · pp. 10–11
3 Methods
The methods evaluate solvated molecules and reactions using static assessments and ML-driven MD, with QM/MM calculations and neural-network training configured for these systems. The HDNNP encodes both QM and MM environments through symmetry functions, including charge-weighted contributions from MM particles.
3.1 Systems
- The study covers solvated molecules and reactions, assessed either retrospectively on sampled configurations or dynamically through MD using ML-provided gradients.Systems include benzene, uracil, retinoic acid, the CH3Cl/Cl− SN2 reaction, and SAM with cytosine.Source: §3.1 · p. 11
3.2 General Computational Details
- QM/MM reference calculations included point charges within Rc in electrostatic embedding, with selected outer solvent atoms retained to avoid bond breaking and artificial charges.Static calculations used a 10−8 eV convergence criterion; MD used 10−6 eV, and long-range MM electrostatics beyond 1.4 nm used a reaction field.Source: §3.2 · p. 11
3.3 Initial Structures
- Reaction starting structures were placed near the transition state and biased by distance restraints, while all solutes were solvated in periodic SPC/E water boxes.The boxes had 1.6 nm minimum solute-wall and 0.25 nm minimum solute-solvent distances; systems were initially relaxed at 0 K.Source: §3.3 · p. 12
3.4 Sampling of Configurations
- After 20,000 steps at 400 K and 1 bar, the first 10,000 were discarded and the remaining frames split 70/20/10 into training, validation, and test sets.The final 1,000 steps served as the external test set; the temporal split was chosen to make test configurations less like training data than a random split would.Source: §3.4 · p. 12
3.5 MD Simulations Using the Fitted ML Models
- Fitted models drove MD at 298 K and 1 bar with a 0.5 fs time step; SAM/cytosine simulations retained the biasing force.The biasing force constant for SAM/cytosine was 2,000 kJ mol−1 nm−2.Source: §3.5 · p. 13
3.6 Neural Networks
- HDNNPs used up to two hidden layers with 10–80 neurons per layer, mila activation, and Adam optimization; training monitored validation loss and selected the lowest-loss model.The gradient-corrected loss L′ was distinguished from the non-gradient-corrected L; both gradient weights defaulted to 1.Source: §3.6 · p. 13
3.7 Symmetry Functions
- QM and MM environments were encoded with symmetry functions; MM-involving functions were partial-charge weighted, and models varied whether they included three-body MM terms.Radial-function cutoffs varied across models, and the symmetry-function parameters were adapted from Ref. 37 to nm and nm−1 units.Source: §3.7 · pp. 13–14, Table 1 · p. 13
4.1 Assessing the Effect of Different Parameters on the Prediction Accuracy
The study tests how gradient correction, delta-learning, training-set size, and interaction cutoffs affect prediction accuracy across benzene, uracil, and a charged transition-state system. It evaluates energies and QM and MM gradients against reference calculations.
4.1.1 Gradient Correction and Many-Body Terms
- For benzene, adding three-body terms alone does not improve accuracy, whereas training on QM and MM gradients regularizes the model and improves energy and gradient predictions.M3's validation energy error falls below chemical accuracy, and increasing its neuron count slightly improves rather than worsens performance.Source: §4.1.1 · pp. 15–16
4.1.2 ∆-Learning
- M7, combining delta-learning with stronger gradient regularization, is the overall best benzene model; delta-learning plus gradient correction improves energies and both gradient predictions even with few neurons.M6's energy MAE is below 4.18 kJ mol−1 on validation but above that threshold on test; M7 improves test performance.Source: §4.1.2 · pp. 17–18, Figure 3 · p. 17
4.1.3 Size of the Training Set
- With only 10% of the benzene training data, M7's energy error remains relatively constant and its gradient performance worsens less than M4's, whose accuracy drops by up to 50%.The comparison uses the same validation and test sets for both training-set sizes.Source: §4.1.3 · p. 18, Table 4 · p. 18
4.1.4 Cutoff Size
- For uracil in water, a 0.52 nm cutoff improves predictions, while extending it to 1.4 nm worsens results on the sparse training set; the same cutoff performs best for CH3Cl/Cl−.The authors suggest larger cutoffs may be desirable with a sufficiently larger training set.Source: §4.1.4 · p. 20, Table 6 · p. 20, Table 7 · p. 21
4.2 Application of ML Models in MD Simulations
The authors tested the delta-learning model in MD simulations of retinoic acid in water and the SAM–cytosine transition state, using a sparse training set. In both systems, trajectories were stable for the simulated durations, and model energies agreed better with DFT than the DFTB baseline.
4.2 Application of ML Models in MD Simulations
- The study used 10,000 training points sampled at 400 K and ran MD at 298 K for two flexible systems; the authors recommend more sophisticated sampling for free-energy profiles.The systems were retinoic acid in water and the SAM–cytosine transition state in water.Source: §4.2 · p. 21, Figure 5 · p. 22
4.2.1 Retinoic Acid in Water
- On retinoic acid’s validation set, delta-learning improved predictions over DFTB for energies, QM gradients, and MM gradients.The comparison uses BP86/def2-TZVP reference data.Source: §4.2.1 · p. 22, Table 8 · p. 22, Figure 6 · p. 23
- MAE = 5.8 kJ mol−1 and RMSE = 7.6 kJ mol−1 for DFTB + ML, versus 52.8 and 54.3 kJ mol−1 for DFTB, over retinoic acid’s 5,000-step trajectory.The DFTB + ML simulation was extended to 200’000 steps; the authors note that more training data may improve practical accuracy for complex molecules.Source: §4.2.1 · pp. 23–24, Figure 7 · p. 24
4.2.2 SAM/Cytosine Transition State in Water
- On the SAM–cytosine transition-state validation set, the delta-learning model clearly outperformed the DFTB baseline.The validation comparison covers energies and QM gradients.Source: §4.2.2 · p. 24, Figure 8 · p. 25, Table 9 · p. 25
- MAE = 18.1 kJ mol−1 and RMSE = 22.2 kJ mol−1 for DFTB + ML, versus 74.9 and 79.4 kJ mol−1 for DFTB, over the SAM–cytosine 2,000-step trajectory.The simulation was extended to 50’000 steps; the authors report stable trajectories for both test systems over their selected durations.Source: §4.2.2 · p. 25, Figure 9 · p. 26
4.3 General Discussion
Delta-learning requires a lower-level calculation at every MD step, but the authors report that DFTB is much cheaper than DFT for the tested SAM–cytosine system. Its use is limited when the lower-level and reference potential-energy surfaces differ substantially.
4.3 General Discussion
- For SAM–cytosine, DFTB took under a second per evaluation versus 60–80 minutes for DFT; delta-learning may be limited when DFTB’s PES differs substantially from the reference and converges slowly or not at all.The authors did not encounter this convergence issue in their presented MD simulations.Source: §4.3 · p. 26
5 Summary and Conclusion
The paper presents an HDNNP workflow for QM/MM MD and evaluates parameter choices and delta-learning. Its delta-learning model produced stable trajectories near DFT accuracy for two larger systems.
5 Summary and Conclusion
- The HDNNP incorporates MM partial charges within a cutoff as an additional element type to describe the QM subsystem in condensed-phase QM/MM simulations.The method targets DFT- or ab initio-level accuracy while avoiding the expense of explicit valence-electron treatment and an SCF procedure at every step.Source: §5 · p. 27
- Gradient correction was crucial for sufficient energy and gradient accuracy in the rigid apolar-solute test, while strong regularization was generally needed; three-body QM–MM terms could be omitted there.The authors qualify that strong regularization may not be necessary with larger training sets.Source: §5 · p. 27
- Delta-learning with DFTB reduced the symmetry-function cutoff for uracil while outperforming classical HDNNP, and stable trajectories near DFT accuracy were obtained for retinoic acid and SAM–cytosine.The authors suggest that improved sampling could further boost performance and identify these findings as promising for practical (QM)ML/MM MD applications.Source: §5 · p. 27