Abstract
This paper investigates shape optimization in mechanical systems characterized by unilateral contact and specific pressure constraints. We compare a high-fidelity gradient-based approach utilizing direct differentiation of the Karush–Kuhn–Tucker (KKT) systems with a derivative-free Bayesian optimization (BO) framework. Through two engineering benchmarks—a wedge-shaped electrical joint and a Marman clamp—we evaluate the trade-offs between mathematical complexity, computational efficiency, and solution accuracy. Our results indicate that while gradient-based methods offer superior precision, BO provides a robust alternative for low-dimensional design spaces without requiring explicit sensitivity analysis.
The Challenge of Contact Optimization
In mechanical engineering, "unilateral contact" refers to bodies that can touch and push against each other but cannot pull (they aren't glued). Optimizing the shape of these bodies is notoriously difficult for two primary reasons. First, the contact region is inherently nonsmooth: as a part's shape evolves, a point might suddenly snap from a non-contacting state to a contacting state. This discontinuity causes spikes in the mathematical objective and constraint functions, creating "cliffs" that traditional optimizers struggle to navigate.
Second, while many studies focus on minimizing overall stress or displacement, few tackle explicit pressure constraints. In real-world applications—such as ensuring a seal does not leak or a high-current joint does not arc—contact pressure must remain within a specific window. This study explores how to navigate these mathematical hurdles to find optimal designs under strict pressure requirements.
Computational Framework
The authors evaluate two distinct philosophies for solving these optimization problems:
- Gradient-Based Optimization: This approach relies on a deep understanding of the underlying physics. It employs the mortar contact method to handle surface interfaces and finite element analysis (FEA) for structural simulation. To optimize, the researchers perform direct differentiation of the Karush–Kuhn–Tucker (KKT) systems. This requires intense mathematical derivation to calculate exactly how infinitesimal changes in shape affect contact pressure. The resulting gradients are fed into the Ipopt solver.
- Bayesian Optimization (BO): This is a "black-box" approach that treats the simulation as an unknown function. Instead of calculating derivatives, BO uses Gaussian Process (GP) surrogate models to estimate the pressure distribution based on previous samples. It utilizes an Expected Constrained Improvement (EIC) acquisition function to decide where to sample next, balancing the exploration of the design space with the exploitation of known feasible regions.
Benchmarking Performance
The methods were tested against two challenging engineering benchmarks:
- The Wedge Problem: Inspired by high-current electrical joints, the goal is to minimize the preload required while ensuring contact pressure remains high enough to prevent electrical arcing across multiple load steps. This problem involved three design variables.
- The Marman Clamp Problem: A complex, axisymmetric model of an industrial pipe clamp where the geometry is defined by cubic Bézier curves. The objective is to maximize system stiffness while keeping contact pressure within a safety range to prevent leaks. This problem involved four design variables.
Key Insights
- Pressure Smoothing: To assist the gradient-based solver, the authors averaged nodal pressures over element segments. This "blurs" the nonsmoothness of the contact interface, allowing the optimizer to find a path to the solution without getting trapped by mathematical noise.
- The "Math vs. Machine" Trade-off: The gradient-based method proved significantly more accurate and found more optimal solutions, but it requires grueling sensitivity derivations specific to every new problem. Conversely, Bayesian Optimization is "plug-and-play" but struggles as the number of design variables increases.
- Feasibility Hurdles: In highly constrained problems like the Marman clamp, the feasible region is extremely small. The authors found that BO requires a feasible starting point to be truly effective, which can be difficult to identify in complex engineering scenarios.
Results and Conclusion
In both benchmarks, the gradient-based method identified optimal designs within 35–48 iterations. Despite having no knowledge of the underlying derivatives, the Bayesian approach found designs very close to the gradient-based optimum within a similar budget of 50 iterations.
The study demonstrates that while contact problems are theoretically nonsmooth, standard solvers like Ipopt can succeed if constraints are formulated carefully (e.g., using segment-averaged pressures). However, scalability remains a limitation; while BO works well for low-dimensional problems (3-4 variables), it is unlikely to scale to topology optimization involving thousands of variables. Future work will focus on developing algorithms that can more effectively "jump" across the discontinuities inherent in contact mechanics.
Summary updated