Uncertainty Quantification (UQ)

UQ rigorously separates two fundamentally distinct sources of error:

• Aleatoric Uncertainty: The irreducible noise inherent in the training data itself, such as thermal fluctuations or instrument measurement error. This uncertainty cannot be reduced by collecting more data.
• Epistemic Uncertainty: The model's ignorance due to limited data or insufficient model capacity. This is the 'reducible' uncertainty that an active learning loop targets, as it is high in unexplored regions of the potential energy surface.

A practical approach to approximating epistemic uncertainty without modifying the core architecture. By training an ensemble of independent neural network potentials with different random initializations, the variance of their predictions for a given atomic configuration serves as a proxy for model confidence.

• High agreement among ensemble members implies low epistemic uncertainty.
• High variance signals an out-of-distribution molecular geometry that requires a new reference calculation via force matching.

A non-parametric Bayesian framework that provides a mathematically rigorous, closed-form posterior distribution for every prediction. Unlike a standard neural network, a GPR model explicitly computes a predictive variance alongside the energy and forces.

• The kernel function encodes physical priors, such as permutation invariance.
• The native uncertainty estimate makes GPR ideal for active learning loops, though it scales poorly to massive datasets compared to deep learning.

A computationally cheap technique to extract uncertainty from a standard deep neural network without retraining. By applying dropout at inference time and performing multiple stochastic forward passes, the variance of the outputs approximates the model's epistemic uncertainty.

• This is formally interpreted as a variational approximation to a deep Gaussian process.
• It provides a lightweight UQ layer for large-scale equivariant neural networks without the cost of full ensembling.

A well-calibrated UQ model ensures that a predicted 90% confidence interval contains the true value exactly 90% of the time. Key diagnostic tools include:

• Reliability Diagrams: Plot expected confidence against observed frequency to detect overconfident or underconfident models.
• Sharpness: Measures the width of the confidence interval. A sharp, calibrated model provides tight, trustworthy error bars, which is critical for assessing the reliability of a free energy perturbation calculation.

The primary operational use-case for UQ in computational chemistry. A 'committee' of models evaluates a vast pool of unlabeled molecular configurations. The configurations with the highest disagreement (maximum epistemic uncertainty) are selected for expensive ab initio molecular dynamics labeling.

• This strategy minimizes the number of costly coupled cluster calculations required to build a globally accurate potential.
• It directly closes the loop between UQ and data acquisition, maximizing the information gained per computation.

The primary consumer of UQ metrics. An iterative training process where the model identifies configurations with high epistemic uncertainty, requests new quantum mechanical calculations for those structures, and retrains. This systematically explores the Potential Energy Surface, building a robust training set without exhaustive sampling. UQ provides the acquisition function that decides what to label next.

A critical decomposition in UQ. Epistemic uncertainty arises from the model's lack of knowledge—regions of chemical space absent from the training data. This is reducible with more data. Aleatoric uncertainty is the inherent, irreducible noise in the data, such as thermal fluctuations or convergence errors in the reference Density Functional Theory calculation. A well-calibrated UQ method distinguishes between the two.

The quality metrics for UQ. Calibration measures whether predicted confidence intervals match observed error frequencies—a 95% confidence interval should contain the true value 95% of the time. Sharpness measures the width of these intervals. A perfect UQ method is both calibrated (honest) and sharp (informative). Poor calibration leads to overconfident failures in Neural Network Potentials.

A practical approach to UQ where multiple independent Neural Network Potentials are trained with different random initializations or on bootstrapped data subsets. The variance of their predictions for a given atomic configuration provides a direct measure of epistemic uncertainty. High variance signals an extrapolatory region. This is computationally expensive but conceptually simple and widely adopted.

A class of machine learning potentials that provide analytical UQ by construction. A Gaussian Process models the potential energy as a distribution of functions, yielding both a mean prediction and a closed-form variance. This variance is a principled measure of model uncertainty. Frameworks like Atomic Cluster Expansion (ACE) can be cast into a Bayesian linear regression form to inherit this property.

A distribution-free, frequentist framework for UQ that provides finite-sample, marginal coverage guarantees. Given a calibration set of atomic configurations, conformal prediction wraps any pre-trained Machine Learning Force Field to output prediction intervals that are guaranteed to contain the true energy or forces with a user-specified probability, without assuming any distribution on the errors.