gomesgroup/sandwich

The mlp-train Python package and general procedure developed by Duarte and co-workers was used to train a reactive machine-learned (https://github.com/duartegroup/mlp-train). A standard active learning (AL) loop using the DSD-PBEP86-D3(BJ)/cc-pVTZ level of theory in ORCA was used. AL used an energy selection method |EMLP − E0| < 1 kcal mol-1. MD simulations initiated at the transition state were used to generate the training set. These simulations were propagated by the Langevin integrator in ASE with initial velocities at 298.15 K sampled from a Maxwell-Boltzmann distribution then propagated using a 0.5 fs timestep. The energies and forces of these configurations were used to train a model that uses the MACE architecture implemented in PyTorch and employing the e3nn library using an NVIDIA A100 80 GB GPU. The MACE training and evaluation codes are distributed via GitHub under the MIT license, available at https://github.com/ACEsuit/mace/. We used a 90:10 training to validation data split, a maximum of 1500 epochs, the weighted energy forces loss used weights of 5 on forces and 1 on energies, and a batch size of 10. The model used two MACE layers, a spherical expansion of lmax = 3, and 4-body messages in each layer (correlation order N = 3), a 256-channel dimension for tensor decomposition, and a radial cutoff of 5 Å. The maximal message equivariance used was L = 2. The irreducible representations of the messages in e3nn notation are 256x0e + 256x1o + 256x2e. All other training parameters were maintained at their default configurations. The AL procedure generated 2905 DFT active learning configurations. The training parameters used achieved chemical accuracy for under various conditions when comparing true and MACE predicated energies and forces for AIMD trajectories. When the MD is initiated from the proton sandwich TS towards to reactant at 398.15 K, MAD = 0.05 kcal mol-1 for energies and MAD = 0.18 kcal mol-1 for forces. When the MD is initiated from the proton sandwich TS towards to product at 398.15 K, MAD = 0.35 kcal mol-1 for energies and MAD = 0.48 kcal mol-1 for forces. Combining MD from both directions of the TS at 398.15 K leads to an overall MAD of 0.20 kcal mol-1 for energies and MAD = 0.33 kcal mol-1 for forces. The AL procedure generated 2905 DFT active learning configurations. The training parameters used achieved chemical accuracy for various conditions when comparing true and MACE predicted energies and forces for AIMD trajectories (see SI for parity plots). The MD was initiated from the proton sandwich transition state in both directions towards the reactant and product at 398.15 K for 500 fs with a 0.5 fs timestep, with energies and forces evaluated every 10 steps. This resulted in a MAD of 0.20 kcal mol-1 for energies and 0.33 kcal mol-1 for forces.