Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
Simulate Time-integrated Coarse-grained Molecular Dynamics with Multi-scale Graph Networks
Authors: Xiang Fu, Tian Xie, Nathan J. Rebello, Bradley Olsen, Tommi S. Jaakkola
TMLR 2023 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The effectiveness of our method is demonstrated in two complex systems: single-chain coarse-grained polymers and multi-component Li-ion polymer electrolytes. For evaluation, we simulate trajectories much longer than the training trajectories for systems with different chemical compositions that the model is not trained on. Structural and dynamical properties can be accurately recovered at several orders of magnitude higher speed than classical force fields by getting out of the femtosecond regime. |
| Researcher Affiliation | Collaboration | Xiang Fu EMAIL Massachusetts Institute of Technology Tian Xie EMAIL Microsoft Research Nathan J. Rebello EMAIL Massachusetts Institute of Technology Bradley D. Olsen EMAIL Massachusetts Institute of Technology Tommi Jaakkola EMAIL Massachusetts Institute of Technology |
| Pseudocode | Yes | Algorithm 1 annealed Langevin dynamics for structural refinement |
| Open Source Code | No | The paper does not explicitly provide a link to its own source code, nor does it state that the code is released or available in supplementary materials for the described methodology. It only references third-party tools like LAMMPS. |
| Open Datasets | Yes | In our first experiment, we focus on simulating single-chain coarse-grained polymers3 within an implicit solvent. We employ the polymers introduced in Webb et al. 2020 We adopt the SPEs introduced in (Xie et al., 2022). |
| Dataset Splits | Yes | We train our model on 100 short class-I MD trajectories (with ten trajectories for validation) of 50k τ, which are not sufficiently long for observable calculations. For evaluation, we use 40 testing class-II polymers using trajectories of 5M τ (100x longer). We train on 530 short MD trajectories of 5 ns (with 30 trajectories for validation) and evaluate on 50 testing SPE trajectories (with distinct polymers unseen during training) of 50 ns long. |
| Hardware Specification | Yes | All models are trained and used for producing long simulations over a single RTX 2080 Ti GPU. |
| Software Dependencies | No | The paper mentions software components like 'Adam optimizer', 'neural networks', and 'LAMMPS', but it does not specify version numbers for any of these to ensure reproducibility. |
| Experiment Setup | Yes | In our implementation, we use 2 hidden layers for all MLPs and 7 message-passing layers for all GNNs. The Embedding GNN GNE has a hidden size of 64, while the Dynamics and Score GNNs have a hidden size of 128. All activation functions in the neural networks are rectified linear units (Re LU). We train the model for 2 million steps. The network is optimized with an Adam optimizer with an initial learning rate of 2 × 10−4, exponentially decayed to 2 × 10−5 over the 2 million training steps. All models are trained and used for producing long simulations over a single RTX 2080 Ti GPU. In the single-chain polymer experiments, the fine-grained R2 g is not computable from the coarse-grained configuration. To obtain the fine-grained R2 g from CG simulations, We first compute the R2 g from the CG configuration and then use an MLP over the latent graph embedding of the dynamics GNN to predict the difference between the CG R2 g and FG R2 g. We then predict the FG R2 g by adding the CG R2 g and the offset predicted by this MLP module. Since the simulation is stable, the refinement module is not used for the single-chain polymers. For the SPE systems, the refinement module is a diffusion model with 20 noise levels ranging from [0.01, 10.]. At simulation time, the refinement is through an annealed Langevin dynamics of 10 steps per noise level and a step size of 5 × 10−5. |