Core Concepts
The author proposes the Gradual Optimization Learning Framework (GOLF) to enhance energy minimization with neural networks, reducing the need for additional data while maintaining accuracy. By enriching training datasets with optimization trajectories, NNPs trained with GOLF outperform traditional methods.
Abstract
The content discusses the challenges in molecular conformation optimization and introduces the GOLF framework to improve neural network training efficiency. Traditional methods rely on physical simulators, but GOLF reduces the need for additional data by utilizing optimization trajectories. The approach is compared to baselines and shows promising results in achieving accurate energy minimization with less computational complexity.
Key points:
- Traditional energy minimization methods are computationally expensive.
- Neural networks can be used to accelerate this process but face challenges due to distribution shift.
- The Gradual Optimization Learning Framework (GOLF) enhances NNPs' performance by incorporating optimization trajectories.
- GOLF reduces the need for additional data while maintaining accuracy in energy minimization tasks.
- Experiments show that NNPs trained with GOLF outperform traditional approaches using significantly less additional data.
Stats
Our results demonstrate that the neural network trained with GOLF performs on par with the oracle on a benchmark of diverse drug-like molecules using significantly less additional data.
We show (see Section 6.2) that NNPs trained with GOLF on the nablaDFT (Khrabrov et al., 2022) perform on par with OG while using 50x less additional data compared to the straightforward approach described in the previous paragraph.