Abstract
Neural surrogates for Partial Differential Equations (PDEs) often suffersignificant performance degradation when evaluated on unseen problemconfigurations, such as novel material types or structural dimensions.Meanwhile, Domain Adaptation (DA) techniques have been widely used in visionand language processing to generalize from limited information about unseenconfigurations. In this work, we address this gap through two focusedcontributions. First, we introduce SIMSHIFT, a novel benchmark dataset andevaluation suite composed of four industrial simulation tasks: hot rolling,sheet metal forming, electric motor design and heatsink design. Second, weextend established domain adaptation methods to state of the art neuralsurrogates and systematically evaluate them. These approaches use parametricdescriptions and ground truth simulations from multiple source configurations,together with only parametric descriptions from target configurations. The goalis to accurately predict target simulations without access to ground truthsimulation data. Extensive experiments on SIMSHIFT highlight the challenges ofout of distribution neural surrogate modeling, demonstrate the potential of DAin simulation, and reveal open problems in achieving robust neural surrogatesunder distribution shifts in industrially relevant scenarios. Our codebase isavailable at https://github.com/psetinek/simshift