PDRL: Descriptor-Based Residual Learning

• PDRL is a post-hoc framework that leverages GNN descriptors to predict residual errors as proxies for uncertainty in interatomic potential predictions.
• It utilizes lightweight auxiliary models to estimate both error norms and deviations, offering a computationally efficient alternative to ensemble methods.
• Empirical results show that PDRL achieves competitive performance in both in-domain and out-of-domain settings, especially for force uncertainty estimation.

to=arxiv_search 公众号天天中彩票 ￣第四色ադրանք 亚历山大发json {"query":"arXiv (Huang et al., 3 Sep 2025) PDRL Post-hoc Descriptor-based Residual Learning for Uncertainty-Aware Machine Learning Potentials", "max_results": 5} to=arxiv_search 全民彩票天天json {"query":"Residual Policy Learning (Silver et al., 2018) residual reinforcement learning demonstrations (Alakuijala et al., 2021)", "max_results": 10} to=arxiv_search 弘鼎 აციას 天天中彩票网络ជាjson {"query":"machine learning interatomic potentials uncertainty quantification ensemble MC-dropout deep kernel learning MACE arXiv", "max_results": 10} Post-hoc Descriptor-based Residual Learning (PDRL) is a post-hoc uncertainty estimation framework for machine learning interatomic potentials (MLIPs) that takes a trained graph neural network potential, extracts its internal descriptors for each atomic environment, computes the model’s residual errors on labeled data, and trains a lightweight auxiliary model to predict those residuals from the descriptors. The method was introduced as a practical alternative to expensive uncertainty quantification methods such as ensembles. Its defining premise is that the discrepancy between MLIP predictions and ground truth values can be modeled directly, and that these learned residuals can serve as proxies for prediction uncertainty (Huang et al., 3 Sep 2025).

1. Definition, scope, and conceptual position

PDRL was proposed in a setting where ensembles are regarded as the gold standard for uncertainty quantification in MLIPs, but their high computational cost can limit practicality. The method is explicitly post-hoc: the base model is already trained and frozen, and uncertainty estimation is added afterward by fitting an auxiliary residual predictor on top of descriptors extracted from that frozen model. The paper presents this as a way to obtain uncertainty estimates without retraining or modifying the original potential, which is especially relevant when a practitioner already has a trained MLIP and wants uncertainty estimates without repeating expensive model development (Huang et al., 3 Sep 2025).

The framework is not presented as a probabilistic model with a rigorous posterior variance. Its stated role is more pragmatic: if predicted residuals correlate with actual errors, they are useful uncertainty proxies. Accordingly, PDRL is described as a residual predictor whose outputs are used as uncertainty signals rather than as a calibrated predictive distribution. The paper is also explicit that it does not perform aleatoric–epistemic decomposition, does not use negative log-likelihood objectives, and does not introduce an explicit calibration transform from predicted residual to confidence interval (Huang et al., 3 Sep 2025).

A common source of confusion is terminological. In broader machine learning, “residual learning” often refers to additive action corrections in reinforcement learning or control. In PDRL, the residual is neither an action correction nor an RL policy; it is the supervised prediction of energy or force error from descriptors of a frozen MLIP. The paper’s own phrasing is that PDRL “leverages the descriptor of a trained graph neural network potential to estimate residual errors” and “models the discrepancy between MLIP predictions and ground truth values, allowing these residuals to act as proxies for prediction uncertainty” (Huang et al., 3 Sep 2025).

2. Formal setting and descriptor extraction

The paper defines a dataset of structures as

$S = \{X_i, E_i, F_i\}_{i=1}^{N},$

where each structure is

$X_i = \{R_i, z_i\},$

with $R_i \in \mathbb{R}^{n_i \times 3}$ denoting atomic positions, $z_i$ atomic numbers, $E_i \in \mathbb{R}$ an energy label, and $F_i \in \mathbb{R}^{n_i \times 3}$ a force label. A trained MLIP defines an energy predictor $f^{\mathrm{energy}}$, with energy prediction $\hat{E}(X) = f^{\mathrm{energy}}(X)$. Force prediction is taken in the usual energy-conserving form,

$$\hat{F}_j(X) = -\frac{\partial f^{\mathrm{energy}}(X)}{\partial r_j}.$$

The residual targets are then the standard discrepancies

$$\Delta E = E - \hat{E}, \qquad \Delta F = F - \hat{F}.$$

At the most basic level, the paper notes that this is the standard residual form $X_i = \{R_i, z_i\},$0 (Huang et al., 3 Sep 2025).

Descriptor extraction is centered on MACE as the base graph neural network potential. The descriptor extractor is denoted

$X_i = \{R_i, z_i\},$1

where $X_i = \{R_i, z_i\},$2 is the number of atoms in the structure and $X_i = \{R_i, z_i\},$3 is the descriptor dimension. For atom $X_i = \{R_i, z_i\},$4, the descriptor is $X_i = \{R_i, z_i\},$5. In the reported experiments, the MACE configuration yields 64 descriptor dimensions. The paper describes these descriptors as the model’s features and uses the official MACE implementation of the descriptor. It does not further formalize a more granular layerwise definition beyond that extraction choice (Huang et al., 3 Sep 2025).

The motivation for descriptor-based residual prediction is that descriptors from a trained GNN should encode local chemical environment, geometric context, element identity and interactions, and information relevant to where the model tends to make errors. This suggests that uncertainty estimation can be improved by learning the map from descriptor to error rather than relying only on descriptor-space rarity or density.

3. Residual-learning variants and objectives

PDRL constructs a post-hoc residual dataset

$X_i = \{R_i, z_i\},$6

On top of this dataset, it proposes two variants for both energy and force: error-norm learning and deviation learning. The first predicts a scalar norm or magnitude of error; the second predicts the residual itself in the same format as the original prediction (Huang et al., 3 Sep 2025).

Variant	Target	Output form
PDRL-norm (energy)	$X_i = \{R_i, z_i\},$7	Structure-level scalar via summed atom-wise outputs
PDRL-diff (energy)	$X_i = \{R_i, z_i\},$8	Structure-level signed residual via summed atom-wise outputs
PDRL-norm (force)	$X_i = \{R_i, z_i\},$9	Atom-wise scalar
PDRL-diff (force)	$R_i \in \mathbb{R}^{n_i \times 3}$0	Atom-wise vector

For energy, the paper avoids a naïve structure-level predictor by using an atom-wise MLP $R_i \in \mathbb{R}^{n_i \times 3}$1 that maps each atomic descriptor to a scalar,

$R_i \in \mathbb{R}^{n_i \times 3}$2

and then sums atom-wise contributions:

$R_i \in \mathbb{R}^{n_i \times 3}$3

This is stated to preserve permutation invariance and scalability to variable-size systems. For energy error-norm learning, the loss is

$R_i \in \mathbb{R}^{n_i \times 3}$4

while energy deviation learning uses

$R_i \in \mathbb{R}^{n_i \times 3}$5

For forces, the paper predicts residuals directly at the atom level. Force error-norm learning uses a scalar atom-wise predictor with target $R_i \in \mathbb{R}^{n_i \times 3}$6,

$R_i \in \mathbb{R}^{n_i \times 3}$7

Force deviation learning uses a vector-valued atom-wise predictor

$R_i \in \mathbb{R}^{n_i \times 3}$8

with loss written compactly in the paper as a squared residual between predicted vector and true force residual; interpreted componentwise, this is the squared Euclidean loss

$R_i \in \mathbb{R}^{n_i \times 3}$9

The uncertainty proxy at inference time is then the predicted residual magnitude or, for vector residuals, naturally the norm of the predicted vector. The paper does not introduce an extra calibration mapping from signed residual prediction to calibrated variance (Huang et al., 3 Sep 2025).

4. Training and inference pipeline

The workflow begins by training a base MLIP. In the experiments, that base model is MACE with the default architecture, 32 channels, radial cutoff $z_i$0, two interaction layers, output descriptor dimension 64, and 100 training epochs. The loss weighting schedule is split into two phases: for the first 75 epochs, the energy weight is 1 and the force weight is 100; for the final 25 epochs, the energy weight is 1000 and the force weight remains 100. After this stage, the base predictor is frozen (Huang et al., 3 Sep 2025).

The second stage extracts descriptors from the frozen model for every labeled structure. The third computes residual targets $z_i$1 and $z_i$2. The fourth stage trains the residual learner. For energy, the residual model is an atom-wise MLP whose outputs are summed over atoms. For force norm, the model emits an atom-wise scalar. For force deviation, it emits an atom-wise vector. The paper reports the following implementation choices:

• PDRL-norm: one hidden layer, ReLU, with softplus right before output.
• PDRL-diff: similar but without softplus.
• Depth: one hidden layer for energy; two hidden layers for forces.
• Optimization: initial learning rate $z_i$3, patience 10 epochs, halve the learning rate when validation error stops improving for 10 epochs, and stop after a maximum of 1000 epochs or when the learning rate reaches $z_i$4.
• Batch sizes: energy PDRL uses batch size 64 structures; force PDRL usually uses batch size 64 atoms; for Ni$z_i$5Al and HME21 force PDRL-diff, batch size 2048 atoms is used (Huang et al., 3 Sep 2025).

At inference time, a test structure is processed by one forward pass through the base MACE model to obtain both prediction and descriptors, followed by one pass through the residual model to obtain the uncertainty proxy. The paper emphasizes the computational consequence: ensemble and MC-dropout require five forward passes of MACE, whereas kNN, GMM, and PDRL require only a single pass, and the extra overhead for PDRL is negligible compared with the MACE forward-pass cost (Huang et al., 3 Sep 2025).

5. Empirical behavior across datasets and uncertainty tasks

The evaluation covers rMD17 subsets (Uracil, Salicylic acid, Malondialdehyde), Ni$z_i$6Al, and HME21. The base architecture is MACE throughout. Baselines are Ensemble, MC-dropout, kNN descriptor distance, and GMM descriptor density. The principal metrics are Spearman correlation between prediction error and uncertainty and AUC for classifying low-error versus high-error examples; in the in-domain AUC benchmark, the lowest 20% error examples are treated as the low-error class and the highest 80% as the high-error class (Huang et al., 3 Sep 2025).

The paper’s headline empirical pattern is asymmetric across target type. For energy uncertainty, PDRL-diff generally performs relatively well. Representative Spearman correlations include $z_i$7 on Uracil, $z_i$8 on Salicylic acid, $z_i$9 on Malondialdehyde, and $E_i \in \mathbb{R}$0 on Ni$E_i \in \mathbb{R}$1Al. On HME21, PDRL-norm slightly exceeds PDRL-diff for energy, with $E_i \in \mathbb{R}$2 versus $E_i \in \mathbb{R}$3 (Huang et al., 3 Sep 2025).

For force uncertainty ranking, PDRL-norm is the strongest variant in-domain and is often highly competitive with or better than all baselines. The reported Spearman correlations are $E_i \in \mathbb{R}$4 on Uracil, $E_i \in \mathbb{R}$5 on Salicylic acid, $E_i \in \mathbb{R}$6 on Malondialdehyde, $E_i \in \mathbb{R}$7 on Ni$E_i \in \mathbb{R}$8Al, and $E_i \in \mathbb{R}$9 on HME21. The HME21 result is especially notable because the ensemble baseline on the same benchmark is $F_i \in \mathbb{R}^{n_i \times 3}$0 (Huang et al., 3 Sep 2025).

The AUC results reinforce this pattern. For force AUC, PDRL-norm reaches 0.85 on Uracil, 0.86 on Salicylic acid, 0.83 on Malondialdehyde, 0.98 on Ni$F_i \in \mathbb{R}^{n_i \times 3}$1Al, and 0.98 on HME21. For energy AUC, the most prominent result is on Ni$F_i \in \mathbb{R}^{n_i \times 3}$2Al, where PDRL-diff reaches 0.90, compared with 0.72 for PDRL-norm and 0.58 for the ensemble baseline (Huang et al., 3 Sep 2025).

The OOD study is performed on Ni$F_i \in \mathbb{R}^{n_i \times 3}$3Al and includes high-temperature OOD, hexagonal phase, cubic phase, and random Ni/Al swaps. For force uncertainty, the overall OOD column reports correlation/AUC values of 0.90/0.99 for Ensemble, 0.82/0.91 for PDRL-norm, and 0.87/0.99 for PDRL-diff. The paper explicitly notes that although PDRL-norm is best for in-domain force uncertainty in Table 1, it is less effective in OOD settings than PDRL-diff, GMM, and kNN. On hexagonal OOD, PDRL-diff obtains correlation 0.89 and AUC 0.97; on swap OOD, it reaches correlation 0.81 and AUC 1.00 (Huang et al., 3 Sep 2025).

The paper interprets these results through the distinction between descriptor geometry and supervised residual learning. Descriptor-only methods such as kNN and GMM can work well when low-density descriptor regions correlate with high error, as in simpler datasets like Ni$F_i \in \mathbb{R}^{n_i \times 3}$4Al. In more complex settings such as HME21, dense descriptor regions can still have elevated force error. PDRL improves in such regimes because it learns the mapping from descriptor to residual directly, rather than assuming that uncertainty is monotonic in descriptor rarity.

6. Strengths, limitations, and relation to broader residual-learning usage

The strongest practical features of PDRL are its post-hoc character, single-pass efficiency, and empirical strength on force uncertainty, especially on complex multi-element data. The method requires neither retraining the base MLIP nor changing its original architecture. It uses the representational content already present in a trained GNN, and the extra predictor is lightweight. The paper therefore positions it as a strong and easy-to-add approximation to ensemble-quality uncertainty ranking, particularly for force predictions (Huang et al., 3 Sep 2025).

Its limitations are equally explicit. PDRL is a residual proxy rather than full probabilistic uncertainty quantification; it does not learn a predictive distribution or variance decomposition. Performance depends on descriptor quality. There is no universally dominant variant: PDRL-diff is usually better for energy uncertainty, PDRL-norm is usually better for in-domain force ranking, and PDRL-diff is better for OOD force detection. The study also focuses on MACE descriptors and does not establish automatic transfer to all MLIP architectures. Finally, because the method is supervised residual learning, it requires labeled structures with ground-truth energies and/or forces to compute residual targets (Huang et al., 3 Sep 2025).

The term “residual learning” here should be distinguished from its usage in reinforcement learning and robotics, where residual methods often mean additive policy or controller corrections of the form “base action + residual action.” That usage is exemplified by Residual Policy Learning and Residual Reinforcement Learning from Demonstrations, where a frozen base controller or cloned policy is improved by an additive action correction (Silver et al., 2018, Alakuijala et al., 2021). PDRL shares the high-level idea of preserving a frozen base model and learning a lightweight correction on top of it, but the correction lives in descriptor-to-error space rather than action space. A plausible implication is that PDRL belongs to a broader family of post-hoc residual methods, while remaining specific in scope to uncertainty-aware MLIPs.

In the paper’s own framing, the methodological priority is residual learning first and uncertainty estimation second. That emphasis is important: PDRL does not estimate uncertainty by ensembling multiple predictors or by introducing stochastic inference, but by modeling the discrepancy between prediction and ground truth directly. Its core claim is therefore not that it reconstructs Bayesian uncertainty, but that descriptor-conditioned residuals can function as accurate, efficient uncertainty proxies for frozen graph neural network potentials (Huang et al., 3 Sep 2025).