TL;DR

• Proposes Structure-Preserving Epistemic Neural Networks (S-PENNs), a general UQ framework for scientific ML models with hard architectural constraints, instantiated for GENERIC dynamics
• S-PENNs attach lightweight epinets to pretrained constrained components, ensuring every sampled realization remains physically admissible by construction without architectural modifications
• Combines S-PENNs with split conformal prediction for post-hoc calibration, yielding prediction intervals with finite-sample marginal coverage guarantees while preserving thermodynamic consistency (first and second laws)
• Validated on three benchmarks: a harmonic oscillator coupled to a heat bath (ODE), an idealized chemical motor (ODE), and a 1D viscoplastic model (PDE), achieving 1–3 orders of magnitude computational cost reduction compared to deep ensembles
• The framework is broadly extensible to scientific ML models in computational mechanics with either hard or soft constraints beyond GENERIC dynamics

Why It Matters

This work addresses a critical gap in scientific machine learning: standard uncertainty quantification methods often violate the physical constraints embedded in structure-preserving models, producing thermodynamically inadmissible predictions. By ensuring UQ respects hard constraints by construction, S-PENNs enable reliable deployment of physics-informed models in safety-critical applications such as computational mechanics and engineering simulation.

Technical Details

• S-PENN Architecture: Lightweight epinets are attached to the constrained components of a pretrained structure-preserving model, rather than modifying the base architecture. Each epinet samples uncertainty while the underlying constrained structure guarantees physical admissibility for every realization.
• GENERIC Dynamics Instantiation: When applied to GENERIC (General Equation for Non-Equilibrium Reversible-Irreversible Coupling) dynamics, the framework produces rollouts that provably preserve both the first and second laws of thermodynamics.
• Conformal Prediction Integration: Split conformal prediction is used as a post-hoc calibration method, providing finite-sample marginal coverage guarantees on prediction intervals without compromising structural constraints.
• Validation Benchmarks: Three numerical examples span both ODE-governed systems (harmonic oscillator + heat bath, idealized chemical motor) and a PDE-governed system (1D viscoplastic model), demonstrating cross-domain applicability.
• Computational Efficiency: S-PENNs reduce computational cost by approximately one to three orders of magnitude compared to deep ensembles, making structure-preserving UQ practical for high-dimensional scientific problems.

Industry Insight

• Structure-preserving ML models are increasingly adopted in computational mechanics and physics-informed applications, but their uncertainty quantification remains a bottleneck; S-PENNs provide a plug-and-play pathway to add rigorous UQ without redesigning existing constrained architectures.
• The combination of hard structural constraints with conformal prediction offers a compelling template for other domains where physical admissibility is non-negotiable, such as climate modeling, structural engineering, and autonomous systems.
• The demonstrated 100–1000x speedup over deep ensembles suggests that structure-preserving UQ can transition from theoretical novelty to practical tooling, particularly for real-time or resource-constrained scientific simulation pipelines.

摘要

提出了结构保持认知神经网络（S-PENNs），这是一种针对具有硬性架构约束的科学机器学习模型的一般性不确定性量化（UQ）框架，并以 GENERIC 动力学为例进行了实例化。
S-PENNs 将轻量级认知网络附加到预训练的约束组件上，确保通过构造保证每个采样实现都保持物理可容许性，而无需修改架构。
将 S-PENNs 与分割共形预测相结合以进行事后校准，从而生成具有有限样本边际覆盖保证的预测区间，同时保持热力学一致性（第一和第二定律）。
在三个基准测试中进行了验证：耦合热浴的谐振子（ODE）、理想化化学电机（ODE）和 1D 粘塑性模型（PDE），与深度集成方法相比，计算成本降低了 1–3 个数量级。
该框架可广泛扩展至计算力学中具有硬性或软性约束的科学机器学习模型，不仅限于 GENERIC 动力学。

深度分析

一句话总结

• 提出了结构保持认知神经网络（S-PENNs），这是一种针对具有硬性架构约束的科学机器学习模型的一般性不确定性量化（UQ）框架，并以 GENERIC 动力学为例进行了实例化。
• S-PENNs 将轻量级认知网络附加到预训练的约束组件上，确保通过构造保证每个采样实现都保持物理可容许性，而无需修改架构。
• 将 S-PENNs 与分割共形预测相结合以进行事后校准，从而生成具有有限样本边际覆盖保证的预测区间，同时保持热力学一致性（第一和第二定律）。
• 在三个基准测试中进行了验证：耦合热浴的谐振子（ODE）、理想化化学电机（ODE）和 1D 粘塑性模型（PDE），与深度集成方法相比，计算成本降低了 1–3 个数量级。
• 该框架可广泛扩展至计算力学中具有硬性或软性约束的科学机器学习模型，不仅限于 GENERIC 动力学。

研究意义

这项工作解决了科学机器学习中的一个关键空白：标准的不确定性量化方法通常会违反结构保持模型中嵌入的物理约束，从而产生热力学上不可容许的预测。通过确保 UQ 通过构造尊重硬性约束，S-PENNs 使得物理信息模型能够在计算力学和工程模拟等安全关键应用中可靠部署。