Temperature-Annealed Boltzmann Generators
Authors: Henrik Schopmans, Pascal Friederich
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We apply this methodology to three molecular systems of increasing complexity and, compared to the baseline, achieve better results in almost all metrics while requiring up to three times fewer target energy evaluations. For the largest system, our approach is the only method that accurately resolves the metastable states of the system. |
| Researcher Affiliation | Academia | 1Institute of Theoretical Informatics, Karlsruhe Institute of Technology, Kaiserstr. 12, 76131 Karlsruhe, Germany 2Institute of Nanotechnology, Karlsruhe Institute of Technology, Kaiserstr. 12, 76131 Karlsruhe, Germany. Correspondence to: Pascal Friederich <EMAIL>. |
| Pseudocode | No | The paper describes the methodology in prose and numbered steps within the text (e.g., 'In one annealing iteration, we perform the following steps: 1. Sample a dataset...', 'Our approach to learn the Boltzmann distribution... can be separated into two phases'), but it does not include explicitly labeled 'Pseudocode' or 'Algorithm' blocks or figures. |
| Open Source Code | Yes | The source code to reproduce our experiments can be found at https://github.com/aimat-lab/TA-BG. |
| Open Datasets | Yes | Furthermore, the ground truth datasets from MD simulations are provided at https://doi.org/10.5281/zenodo.15526429. |
| Dataset Splits | Yes | The ground truth test datasets at 300 K contain 1 107 samples, the datasets used for the forward KLD experiments contain 1 106 samples. The additional validation datasets at 300 K and 1200 K contain 1 106 samples. |
| Hardware Specification | Yes | To determine the wall times, we ran 4 experiments in parallel on a compute node with 2 AMD EPYC Rome 7402 CPU (24 cores each) and 4 NVIDIA A100 GPU. |
| Software Dependencies | Yes | All ground truth simulations have been performed with Open MM 8.0.0 (Eastman et al., 2024)... To implement the normalizing flow models and the internal coordinate representations, we used the bgflow (No e, 2024) and nflows (Conor Durkan et al., 2020) libraries with Py Torch (Paszke et al., 2019). |
| Experiment Setup | Yes | Table 5. Hyperparameters of the TA-BG experiments, annealing from 1200 K to 300 K. The cosine annealing learning rate scheduler is applied within each annealing iteration, so the learning rate resets in the beginning of each new annealing iteration. Annealing iteration here also refers to the fine-tuning iterations with Ti+1 = Ti. DIPEPTIDE TETRAPEPTIDE HEXAPEPTIDE GRADIENT DESCENT STEPS PER ANNEALING ITERATION 30 000 20 000 20 000 LEARNING RATE 5 10 6 1 10 5 5 10 6 BATCH SIZE 2048 4096 2048 LR SCHEDULER COSINE ANNEALING BUFFER SAMPLES DRAWN PER ANNEALING ITERATION 5 106 5 106 1 107 BUFFER RESAMPLED TO 2 106 2 106 2 106 |