Lifelong Machine Learning Potentials

Marco Eckhoff, Markus Reiher

TL;DR

This paper introduces "lifelong machine learning potentials" (lMLPs) that can continuously learn and adapt to new chemical systems without forgetting previous knowledge. They achieve this by combining novel element-embracing descriptors (eeACSFs), a robust training optimizer (CoRe), and uncertainty quantification, paving the way for more versatile and efficient atomistic simulations.

What Problem Does This Paper Address?

Accurate prediction of molecular behavior relies on understanding their potential energy surfaces (PESs). While high-fidelity methods like Density Functional Theory (DFT) provide accurate PESs, they are computationally expensive for large-scale simulations. Machine learning potentials (MLPs) offer a solution by learning accurate PESs from DFT data, providing near-force-field computational cost with DFT accuracy. However, traditional MLPs have a significant drawback: they are typically trained for a specific system and require retraining from scratch for new systems or to incorporate new data, leading to "catastrophic forgetting" of previously learned knowledge. Furthermore, many common MLP descriptors struggle to efficiently represent systems with a large number of different chemical elements, leading to a combinatorial explosion in descriptor complexity. This paper aims to overcome these limitations by developing MLPs that can learn continuously and handle diverse chemical elements efficiently.

Methods

The authors introduce several key innovations:

  • Element-Embracing Atom-Centered Symmetry Functions (eeACSFs): These are novel structural descriptors that incorporate information from the periodic table (element's period, group, and block) alongside traditional atomic environment features. This design makes the descriptor size independent of the number of chemical elements, addressing the scaling issue of traditional descriptors.
  • Lifelong Machine Learning Potential (lMLP) Concept: This framework enables MLPs to learn incrementally from a continuous stream of new data. It leverages uncertainty quantification to ensure a predefined level of accuracy is maintained and employs continual learning strategies to prevent catastrophic forgetting.
  • Continual Resilient (CoRe) Optimizer: This is a novel first-order gradient-based optimizer designed for stochastic iterative optimizations. It combines the robustness of RPROP with the performance of Adam, incorporating adaptive learning rates, plasticity factors to mitigate forgetting, and weight decay to improve convergence and accuracy.
  • Lifelong Adaptive Data Selection: This strategy dynamically selects training data, excluding redundant or incorrect data points while retaining important information through rehearsal. This process helps manage the growing data size in lifelong learning scenarios and improves training efficiency.
  • Uncertainty Quantification: The authors utilize ensemble methods to estimate prediction uncertainties. This allows the lMLP to signal when it's operating outside its reliable prediction range, indicating the need for more data or further training.

Datasets

The study utilizes several datasets for training and evaluation:

  • Reference Data Set A: Contains 2026 conformations for ten different gas-phase SN2 reactions involving four chemical elements. This dataset was used to compare the performance of ACSF and eeACSF representations.
  • Reference Data Set B: A larger dataset with 8600 conformations for 42 different SN2 reactions, involving ten chemical elements. This dataset was crucial for testing the eeACSF representation and the lifelong learning capabilities.
  • Reference Data Set C: Constructed by adding structurally distorted conformations to dataset B, totaling 21151 conformations. This dataset was used to evaluate the lMLP's ability to handle expanded conformation spaces.
  • Validation Set: A dense grid of conformations for the Br⁻ + CH₃Cl reaction, specifically designed to test the smoothness and accuracy of the learned potential energy surface and the uncertainty quantification outside the training data distribution.

Key Insights and Ideas

  • Element Independence is Key: The eeACSF descriptor's ability to represent diverse chemical elements without a combinatorial explosion in size is a significant breakthrough for developing general-purpose MLPs.
  • Lifelong Learning is Feasible: The paper demonstrates that MLPs can indeed learn continuously without catastrophic forgetting, mimicking how humans learn. This opens doors for dynamic model updates and adaptation.
  • Uncertainty is Your Friend: Uncertainty quantification is not just a measure of error but a crucial component for guiding the lifelong learning process, indicating when and where the model needs more information.
  • CoRe Optimizer is a Game Changer: The CoRe optimizer significantly outperforms standard optimizers like SGD, RPROP, and Adam in terms of both speed and final accuracy for MLP training, especially in the context of lifelong learning.
  • Data Selection Matters: Intelligent selection and management of training data are vital for efficient lifelong learning, preventing models from being overwhelmed by large datasets and ensuring focus on relevant information.

Key Results

  • eeACSFs vs. ACSFs: For systems with up to four elements, ACSF and eeACSF representations show similar accuracy. However, for systems with more elements, eeACSFs become computationally advantageous due to their fixed size.
  • Ensemble Accuracy: The ensemble of HDNNPs trained with eeACSFs achieved state-of-the-art accuracy for SN2 reactions involving ten elements, comparable to MLPs trained on fewer elements.
  • CoRe Optimizer Performance: The CoRe optimizer consistently outperformed SGD, RPROP, and Adam, achieving significantly lower RMSE values for both energies and forces. For instance, on dataset B, CoRe achieved an ensemble RMSE(E) of 2.6 meV atom⁻¹ and RMSE(F) of 64 meV Å⁻¹.
  • Lifelong Learning Benefits: The lMLP demonstrated comparable accuracy to training on a stationary batch of all data, even when new data was introduced late in the training process or when expanding the conformation space.
  • Uncertainty Quantification Reliability: The uncertainty quantification was found to be equal to or larger than the absolute error for over 98% of the validation data, reliably indicating prediction accuracy.

Limitations and Future Work

  • Model Architecture Limitations: The current work uses a second-generation HDNNP architecture. Future work could explore extending the lMLP concept to more advanced architectures, such as those incorporating explicit long-range interactions or graph-based representations.
  • Data Consistency: Current MLPs require consistent training data (e.g., calculated with the same method and basis set). Future research could focus on developing methods to handle data from different reference methods or electronic states within a single lMLP.
  • Generalization to Diverse Systems: While the current work focuses on SN2 reactions, future studies need to explore a wider range of chemical systems and reaction types to fully validate the universality of the lMLP concept.
  • Autonomous Workflows: The authors suggest the need for autonomous workflows and user-friendly interfaces to facilitate the development and application of lMLPs.
  • Continual Expansion of Model Architectures: To avoid information capacity limitations, future work should investigate algorithms for continually expanding model architectures as new data and complexity are introduced.

Summary updated

Abstract

Machine learning potentials (MLPs) trained on accurate quantum chemical data can retain the high accuracy, while inflicting little computational demands. On the downside, they need to be trained for each individual system. In recent years, a vast number of MLPs have been trained from scratch because learning additional data typically requires retraining on all data to not forget previously acquired knowledge. Additionally, most common structural descriptors of MLPs cannot represent efficiently a large number of different chemical elements. In this work, we tackle these problems by introducing element-embracing atom-centered symmetry functions (eeACSFs), which combine structural properties and element information from the periodic table. These eeACSFs are key for our development of a lifelong machine learning potential (lMLP). Uncertainty quantification can be exploited to transgress a fixed, pretrained MLP to arrive at a continuously adapting lMLP, because a predefined level of accuracy can be ensured. To extend the applicability of an lMLP to new systems, we apply continual learning strategies to enable autonomous and on-the-fly training on a continuous stream of new data. For the training of deep neural networks, we propose the continual resilient (CoRe) optimizer and incremental learning strategies relying on rehearsal of data, regularization of parameters, and the architecture of the model.
BibTeX
@article{eckhoff2023lifelong,
  author = {Marco Eckhoff and Markus Reiher},
  title = {{Lifelong Machine Learning Potentials}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2023},
  doi = {10.1021/acs.jctc.3c00279},
}
Figure