Darzin International Company Multidisciplinary Engineering Science Open 2 Serial Number 2 2025 01 01 https://jmesopen.com/index.php/jmesopen/workflow/index/38/5#publication/contributors https://jmesopen.com/index.php/jmesopen/workflow/index/38/5#publication/contributors 1 11 EN Journal Article 2024 10 20 <p>This study aimed to synthesize and critically evaluate the role of differentiable physics frameworks in advancing multidisciplinary design optimization (MDO), focusing on implicit gradient computation, adjoint-based sensitivity analysis, and robustness of optimization performance under distributional shifts. A qualitative review design was adopted to examine sixteen peer-reviewed articles published between 2015 and 2025 that addressed differentiable physics, adjoint methods, and robust optimization in MDO contexts. Data collection relied exclusively on systematic literature review procedures across Scopus, Web of Science, IEEE Xplore, and ScienceDirect databases. Studies were selected through purposive sampling until theoretical saturation was achieved. Data were analyzed thematically using Nvivo 14 software through open, axial, and selective coding stages. Emergent concepts were organized into four major themes: differentiable physics foundations, adjoint-based optimization, robustness under distribution shift, and future integration challenges. The synthesis revealed that implicit differentiation and adjoint-based gradient computation form the computational backbone of differentiable physics in MDO, enabling scalable and memory-efficient sensitivity analysis across coupled physical domains. However, computational efficiency, gradient stability, and numerical conditioning remain significant challenges that limit generalization across problem types. The findings also indicate that while differentiable frameworks have achieved theoretical maturity, their robustness under distributional shift—such as environmental or boundary condition changes—remains underexplored. Integration with uncertainty quantification, Bayesian robustness, and domain adaptation is emerging as a promising solution. Additionally, the analysis underscored the lack of standardized benchmarks and reproducibility protocols, which constrains cross-study validation. Differentiable physics represents a paradigm shift in engineering optimization by bridging first-principles simulation and gradient-based learning. Yet, realizing its full potential requires methodological advancements in implicit solvers, cross-domain adjoint coupling, and robustness-aware design. Future work should emphasize scalable algorithms, reproducible benchmarking, and integration with real-world uncertainty modeling to foster reliable and interpretable differentiable MDO systems.</p> https://jmesopen.com/index.php/jmesopen/article/download/39/34