Abstract:Statistical learning algorithms provide a generally-applicable framework to sidestep time-consuming experiments, or accurate physics-based modeling, but they introduce a further source of error on top of the intrinsic limitations of the experimental or theoretical setup. Uncertainty estimation is essential to quantify this error, and make application of data-centric approaches more trustworthy. To ensure that uncertainty quantification is used widely, one should aim for algorithms that are reasonably accurate, but also easy to implement and apply. In particular, including uncertainty quantification on top of an existing architecture should be straightforward, and add minimal computational overhead. Furthermore, it should be easy to manipulate or combine multiple machine-learning predictions, propagating uncertainty over further modeling steps. We compare several well-established uncertainty quantification frameworks against these requirements, and propose a practical approach, which we dub direct propagation of shallow ensembles, that provides a good compromise between ease of use and accuracy. We present benchmarks for generic datasets, and an in-depth study of applications to the field of atomistic machine learning for chemistry and materials. These examples underscore the importance of using a formulation that allows propagating errors without making strong assumptions on the correlations between different predictions of the model.

What problem does this paper attempt to address?

### What problems does this paper attempt to solve? This paper aims to solve the problem of how to effectively quantify and propagate uncertainty when using machine - learning (ML) models for prediction. Specifically, the author focuses on the uncertainty quantification (UQ) of machine - learning interatomic potentials (MLIPs) used in atomic - scale simulations. The following are the key problems that the paper attempts to solve: 1. **Introducing new sources of error**: - Although ML models can replace time - consuming experiments or physics - based modeling, they themselves introduce additional sources of error. These errors may stem from the limited sampling of training data, the choice of model architecture, etc. 2. **Increasing the trustworthiness of data - driven methods**: - Uncertainty estimation is crucial for quantifying these errors, making data - driven methods more reliable. By providing reliable uncertainty estimates, the trust in model predictions can be enhanced. 3. **Simplifying the implementation of uncertainty quantification**: - Existing uncertainty quantification methods are often complex and computationally expensive. In order to make uncertainty quantification widely applicable, methods that are easy to implement and have low computational cost need to be developed. 4. **Combining multiple machine - learning predictions**: - In practical applications, it is often necessary to combine multiple machine - learning predictions to evaluate the uncertainty of the final result. For example, in molecular dynamics simulations, the model needs to be called multiple times and the uncertainty needs to be propagated. 5. **Ensuring the accuracy of uncertainty propagation**: - When combining multiple predictions or further processing the model output, ensure that the uncertainty can be accurately propagated to derived quantities or averages. This is especially important for tasks such as evaluating thermodynamic properties. ### Proposed solutions To solve the above problems, the author proposes the "Direct Propagation of Shallow Ensembles" (DPOSE) method. This method has the following advantages: - **Easy to implement**: It can be easily added on top of existing regression models. - **Low computational cost**: It does not significantly increase the computational burden. - **High accuracy**: It can effectively propagate uncertainty while maintaining accuracy. - **Robust**: It is applicable to various machine - learning models and can handle the correlations between different predictions. ### Specific application scenarios The paper verifies the effectiveness of the DPOSE method through a series of benchmark tests and practical applications, including: - **Benchmark tests on general datasets**: Evaluate the performance of different uncertainty quantification methods in general regression tasks. - **Molecular formation energy prediction**: Demonstrate the performance of DPOSE in predicting molecular formation energy. - **Condensed - matter system simulations**: Applied to the simulations of liquid water, barium titanate, and solid - state electrolyte Li₃PS₄ to evaluate its performance in potential energy and force prediction. - **QM9 molecular dataset**: Further verify the effectiveness of DPOSE in molecular property prediction. Through these experiments, the author proves that the DPOSE method provides a reliable and efficient uncertainty quantification tool in atomic - scale simulations.

Uncertainty quantification by direct propagation of shallow ensembles

Probabilistic Computation with Emerging Covariance: Towards Efficient Uncertainty Quantification

Uncertainty Quantification and Propagation in Atomistic Machine Learning

Uncertainty quantification in classical molecular dynamics

Methods for comparing uncertainty quantifications for material property predictions

Single-model uncertainty quantification in neural network potentials does not consistently outperform model ensembles

Quantifying Epistemic Uncertainty in Deep Learning

Deeply Uncertain: Comparing Methods of Uncertainty Quantification in Deep Learning Algorithms

Evaluating Scalable Uncertainty Estimation Methods for Deep Learning-Based Molecular Property Prediction

Uncertainty Quantification and Propagation in Surrogate-based Bayesian Inference

Statistical methods for resolving poor uncertainty quantification in machine learning interatomic potentials

Robust and scalable uncertainty estimation with conformal prediction for machine-learned interatomic potentials

Uncertainty Quantification via Stable Distribution Propagation

Ensembles Are Required to Handle Aleatoric and Parametric Uncertainty in Molecular Dynamics Simulation

Scalable Bayesian Uncertainty Quantification for Neural Network Potentials: Promise and Pitfalls

Uncertainty Quantification for Deep Learning

Fast and Accurate Uncertainty Estimation in Chemical Machine Learning

Spatially resolved uncertainties for machine learning potentials

Explainable uncertainty quantifications for deep learning-based molecular property prediction

Uncertainty Quantification and Propagation in Computational Materials Science and Simulation-Assisted Materials Design

Deep Ensembles vs. Committees for Uncertainty Estimation in Neural-Network Force Fields: Comparison and Application to Active Learning