Less is more: sampling chemical space with active learning (1801.09319v2)

Abstract: The development of accurate and transferable ML potentials for predicting molecular energetics is a challenging task. The process of data generation to train such ML potentials is a task neither well understood nor researched in detail. In this work, we present a fully automated approach for the generation of datasets with the intent of training universal ML potentials. It is based on the concept of active learning (AL) via Query by Committee (QBC), which uses the disagreement between an ensemble of ML potentials to infer the reliability of the ensemble's prediction. QBC allows the presented AL algorithm to automatically sample regions of chemical space where the ML potential fails to accurately predict the potential energy. AL improves the overall fitness of ANAKIN-ME (ANI) deep learning potentials in rigorous test cases by mitigating human biases in deciding what new training data to use. AL also reduces the training set size to a fraction of the data required when using naive random sampling techniques. To provide validation of our AL approach we develop the COMP6 benchmark (publicly available on GitHub), which contains a diverse set of organic molecules. Through the AL process, it is shown that the AL-based potentials perform as well as the ANI-1 potential on COMP6 with only 10% of the data, and vastly outperforms ANI-1 with 25% the amount of data. Finally, we show that our proposed AL technique develops a universal ANI potential (ANI-1x) that provides accurate energy and force predictions on the entire COMP6 benchmark. This universal ML potential achieves a level of accuracy on par with the best ML potentials for single molecule or materials, while remaining applicable to the general class of organic molecules comprised of the elements CHNO.

Active Learning for Sampling Chemical Space in ML Potentials

The paper discusses a novel methodology for developing ML models that are capable of predicting molecular energetics with greater efficiency. It introduces an automated approach to dataset generation using active learning (AL) through a Query by Committee (QBC) strategy. This method capitalizes on disagreement among a committee of models to pinpoint areas of chemical space with higher predictive error, enabling targeted data sampling.

Methodological Advancements

The paper implements a two-component strategy for enhancing ML potential training:

1. Dataset Reduction: The algorithm identifies and eliminates redundancies in existing data, maintaining predictive performance while minimizing dataset size. This optimization significantly reduces computational resource requirements.
2. Active Learning via QBC: AL using QBC selects new training data by analyzing where model predictions diverge. This is done through a statistical framework that evaluates prediction variance among different models within an ensemble. The process iterates over configurational and conformational sampling, improving model accuracy with only a fraction of the data typically required.

The paper introduces the Comprehensive Machine-learning Potential Benchmark (COMP6) suite to validate model performance. COMP6 includes datasets of varying size and complexity, ensuring robust testing of model extensibility and transferability across organic molecules.

Key Findings

• Efficient Data Utilization: The AL-based approach demonstrates that the same accuracy as the ANI-1 potential can be achieved with only 10% of the original data. It notably outperforms ANI-1 with just 25% of the dataset size.
• Training a Universal Potential: The developed ANI-1x potential, a product of this AL technique, achieves similar accuracy to specific QM methods for molecular systems containing hydrogen, carbon, nitrogen, and oxygen.
• Improved Prediction Errors: With comprehensive validation on the COMP6 suite, the ANI-1x outperforms prior models in all evaluated metrics, including energy and force prediction errors.

Implications and Future Outlook

The development of ANI-1x illustrates that active learning can drastically improve the efficiency of data utilization in training ML potentials. This approach minimizes the traditionally required extensive QM data, allowing for faster development of universal potentials that are more broadly applicable.

Practically, these advancements could accelerate the simulation of molecular systems in computational chemistry, facilitating drug design, material discovery, and beyond. Theoretically, it lays the groundwork for more adaptive and responsive AI models in chemistry, which can predict interactions in previously unexplored chemical spaces with minimal data.

Conclusion

This research marks a significant contribution to the optimization of ML models in computational chemistry through innovative use of active learning. While the work underscores the potential for universal model applications, it also suggests a pathway forward for the development of ML potentials that require less data but offer high accuracy. Future research should continue to explore and refine these strategies, potentially extending applications to other domains such as materials science.