Skip to content

A Multiple-Fidelity Method for Accurate Simulation of MoS2 Properties Using JAX-ReaxFF and Neural Network Potentials

Summary

The authors combine JAX-ReaxFF (automatic-differentiation ReaxFF optimization against DFT) with neural network potentials: ReaxFF-generated data pretrain SchNet, then DFT finetunes ("multiple fidelity"), improving MoS2 property modeling; a Mo-S-H extension with MACE reports ~20% lower energy RMSE versus training from scratch.

Methods

Step A: Fit ReaxFF for MoS2 with JAX-ReaxFF using DFT training items from prior work (cited ref 17 in the letter): 372 total training items split into 3 charge-based, 78 geometry-based, and 291 energy-based items over 425 geometries (D1), including 88 Mo-S-only energy items on 109 geometries (D2). Loss is weighted MSE between ReaxFF and DFT targets (weights from ref 17). Step B: Train NNP-SchNet on 10,000 MD samples from JAX-ReaxFF-MoS2. Step C: Finetune on 400 held-out DFT points (D2) without new DFT. Mechanical tests: LAMMPS stress-strain and stress-fluctuation elastic properties (details in SI Section S2). NEB in LAMMPS with JAX-ReaxFF-MoS2 for 1T to 2H barriers (Table 4, comparison to literature ~1.04 eV per formula unit from Guo et al.). Mo-S-H: NNP-MACE pretrained on ReaxFF data then finetuned.

Software/autodiff context. JAX-ReaxFF uses automatic differentiation through the ReaxFF energy expression so gradients with respect to parameters accelerate weighted least-squares optimization versus finite-difference parameter sweeps. SchNet training consumes MD snapshots from the optimized ReaxFF field to build local environment embeddings before DFT finetuning on held-out energies/forces.

1 — MD application (property / stress runs). Engine: LAMMPS (stress-strain, NEB paths cited) with JAX-ReaxFF-MoS2 as stated. System: MoS2 supercells / unit cells in 3D PBC for the mechanical and kinetics production work (~400+ atom-scale cells in letter/SI, exact counts: see PDF). NVT Molecular dynamics to generate the 10,000-snapshot ReaxFF-trained trajectory pool; 1 ns-class equilibration and production duration in SI S2 (if not in this blurb, treat as N/A here and read SI). Thermostat, timestep, barostat, GPa stress targets: 0 GPa isotropic hydrostatic not assumed—use NPT only where the letter specifies (otherwise N/A in this one-page summary). T-controlled sampling as in the letter. Electric field: N/A. Replica / umbrella / metadynamics: N/A; NEB is the barrier tool. 2 — Force-field training (JAX-ReaxFF, SchNet, MACE). The steps A–C in Methods above. 3 — Static QM (DFT training). DFT from reference 17 and held-out D2; M06-2X etc. not the headline—see cited training literature for the exact functional in each training item.

Findings

JAX-ReaxFF-MoS2 yields tensile strength ~22 GPa, closer to experiment and DFT (~21-22 GPa) than original ReaxFF (~24 GPa). Elastic properties and Poisson ratio show reduced overestimation versus legacy ReaxFF (Table 3, SI Table S2). NEB paths for SL MoS2 1T-2H match DFT reference barriers better in the transition region than original ReaxFF. SchNet with ReaxFF pretraining plus DFT finetune improves thermodynamic descriptors (e.g. convex hull, sulfur vacancy formation, S8 interaction tests in the letter). MACE on Mo-S-H reports ~20% lower energy RMSE versus from-scratch training. The authors state the pipeline generalizes as transfer learning from a high-throughput cheap model to a data-sparse accurate model. Comparisons to experiment/literature (Table 3, Guo et al., bulk moduli) are in the main bullets; sensitivity to data splits and pretraining is discussed in the letter/ SI rather than re-derived here. Use the J. Phys. Chem. Lett. PDF and SI for numerical authority.

Limitations

JAX-ReaxFF, SchNet, and MACE inherit training-set and functional biases; the multiple-fidelity pipeline is demonstrated for MoS₂ and Mo–S–H extensions—transfer to other TMDs or electrode environments requires refitting and validation.

Relevance to group

Illustrates transfer learning from ReaxFF-generated data to neural potentials for 2D chalcogenides—adjacent to ReaxFF and MLIP methodology threads in the corpus.

Citations and evidence anchors