SIMSHIFT: A Benchmark for Adapting Neural Surrogates to Distribution Shifts

TL;DR

SIMSHIFT provides a robust benchmark to study how neural surrogates for PDEs cope with distribution shifts in industrial settings by coupling four realistic tasks with a formal UDA framework. The framework jointly optimizes a conditioning network and a neural surrogate under reconstruction and domain-adaptation losses, using density-ratio based model selection to navigate unlabeled target domains. Across tasks, UDA improves target performance relative to unregularized baselines, though gains vary by metric and dataset, and unsupervised model selection proves as impactful as the adaptation method. The work highlights both the potential and the gaps in achieving robust neural surrogates under distribution shifts, and it lays the groundwork for physics-informed UDA and broader multi-parameter shifts in industrial design problems.

Abstract

Neural surrogates for Partial Differential Equations (PDEs) often suffer significant performance degradation when evaluated on problem configurations outside their training distribution, such as new initial conditions or structural dimensions. While Unsupervised Domain Adaptation (UDA) techniques have been widely used in vision and language to generalize across domains without additional labeled data, their application to complex engineering simulations remains largely unexplored. In this work, we address this gap through two focused contributions. First, we introduce SIMSHIFT, a novel benchmark dataset and evaluation suite composed of four industrial simulation tasks spanning diverse processes and physics: hot rolling, sheet metal forming, electric motor design and heatsink design. Second, we extend established UDA methods to state-of-the-art neural surrogates and systematically evaluate them. Extensive experiments on SIMSHIFT highlight the challenges of out-of-distribution neural surrogate modeling, demonstrate the potential of UDA in simulation, and reveal open problems in achieving robust neural surrogates under distribution shifts in industrially relevant scenarios. Our codebase is available at https://github.com/psetinek/simshift

Figures (26)

• Figure 1: Schematic overview of the SIMSHIFT framework. In training, the model has access to inputs (e.g., parameters and meshes) with the corresponding outputs $(x, y)$ from the source domain (left, blue), while only inputs $x'$ from the target domain (right, yellow) are available. The neural operator $g$ and the conditioning network $\phi$ are shared across domains and jointly optimized. Two loss terms are used: $\mathcal{L}_{\text{recon}}$, computed on source labels, and $\mathcal{L}_{\text{DA}}$, which aligns source and target feature representations produced by $\phi$. After training, unsupervised model selection strategies choose $\theta_{k1}$, which is expected to perform best on the target domain.
• Figure 2: Overview of the hot rolling (top) and sheet metal forming (bottom) simulation scenarios.
• Figure 3: Overview of electric motor design (left) and heatsink design (right) simulation scenarios.
• Figure 4: SIMSHIFT's predefined distribution shifts. $N$ denotes the number of samples in the respective domain.
• Figure 5: Target error scaling with increasing domain gap. We show the averaged nRMSE across all fields for the easy, medium, and hard shifts on the hot rolling task. We compare PointNet models without UDA to DARE-GRAM combined with IWV and TB selection (lower bound). Error bars indicate the standard deviation across four seeds. Furthermore, we highlight potentials of UDA algorithm and model selection improvements.