Robust design optimization for a nonlinear system via Bayesian neural network enhanced polynomial dimensional decomposition

Uncertainties such as manufacturing tolerances cause performance variations in complex engineering systems, making robust design optimization (RDO) essential. However, simulation-based RDO faces high computational cost for statistical moment estimation, and strong nonlinearity limits the accuracy of conventional surrogate models. This study proposes a novel RDO method that integrates Bayesian neural networks (BNN) with polynomial dimensional decomposition (PDD). The method employs uncertainty-based active learning to enhance BNN surrogate accuracy and a multi-point single-step strategy that partitions the design space into dynamically adjusted subregions, within which PDD analytically estimates statistical moments from BNN predictions. Validation through a mathematical benchmark and an electric motor shape optimization demonstrates that the method converges to robust optimal solutions with significantly fewer function evaluations. In the ten-dimensional benchmark, the proposed method achieved a 99.97% mean reduction, while Gaussian process-based and Monte Carlo approaches failed to locate the global optimum. In the motor design problem, the method reduced cogging torque by 94.75% with only 6644 finite element evaluations, confirming its computational efficiency for high-dimensional, strongly nonlinear engineering problems.

💡 Research Summary

This paper introduces a novel robust design optimization (RDO) framework that synergistically combines Bayesian neural networks (BNNs) with polynomial dimensional decomposition (PDD). The motivation stems from the high computational cost of simulation‑based RDO, especially when estimating statistical moments (mean and variance) of performance functions under aleatory uncertainties such as manufacturing tolerances. Conventional surrogate models—polynomial chaos expansions, Kriging, support‑vector machines, or deterministic neural networks—struggle with strongly nonlinear responses and high‑dimensional input spaces.

Methodology
The authors first employ a BNN as a probabilistic surrogate. Unlike deterministic networks, a BNN treats weights and biases as random variables, enabling explicit quantification of epistemic uncertainty. Training is performed via variational inference, maximizing the evidence lower bound (ELBO) with mini‑batch stochastic gradient descent, which scales linearly with data size and remains tractable for high‑dimensional problems.

An uncertainty‑driven active learning loop is built on top of the BNN. At each iteration, the predictive variance of the BNN is examined; regions with high variance are selected for additional high‑fidelity evaluations (finite‑element simulations). The newly acquired data are added to the training set, and the BNN is retrained, thereby concentrating samples where the model is least certain and dramatically improving data efficiency.

To compute statistical moments analytically, the authors embed the BNN within a polynomial dimensional decomposition (PDD) scheme. PDD decomposes a multivariate function into a hierarchy of low‑order component functions (ANOVA‑style). By fitting the PDD coefficients through least‑squares regression on the BNN‑predicted responses, the mean and variance of the surrogate can be obtained in closed form: the mean equals the first coefficient, and the variance is the sum of squares of all higher‑order coefficients. This eliminates the need for costly Monte‑Carlo sampling once the BNN‑PDD surrogate is built.

The framework also adopts a multi‑point single‑step (MPSS) strategy for handling the design space. The design domain is partitioned into dynamically adjusted sub‑regions. Within each sub‑region, a local RDO problem is solved using the BNN‑PDD surrogate, while a global search component (e.g., evolutionary algorithm) explores the sub‑region to avoid reliance on potentially unreliable gradients in highly nonlinear zones. After each iteration, sub‑region boundaries are resized based on progress, preventing premature convergence to local optima and enhancing robustness.

Algorithmic Flow

1. Initialize design variables and uncertainty distributions; generate an initial set of high‑fidelity samples.
2. Train the BNN on these samples.
3. Perform active learning: evaluate predictive variance, select new points, run high‑fidelity simulations, augment the dataset, and retrain the BNN.
4. Partition the design space (MPSS) and, for each sub‑region, compute PDD coefficients from the BNN predictions.
5. Analytically evaluate mean, variance, and the weighted‑sum objective; enforce constraints.
6. Conduct local optimization within each sub‑region; adjust sub‑region sizes.
7. Iterate steps 3‑6 until convergence or budget exhaustion.

Validation
Two case studies demonstrate the method’s effectiveness:

10‑dimensional benchmark: A synthetic nonlinear function with ten independent random inputs was optimized. The proposed BNN‑PDD approach achieved a 99.97 % reduction in the objective mean, locating the global optimum with far fewer evaluations than Gaussian‑process (GP) surrogates or plain Monte‑Carlo (MC) methods, which failed to converge to the optimum.

Electric motor cogging‑torque reduction: An eight‑design‑variable motor model with two aleatory uncertainties (material properties and manufacturing tolerances) was optimized to minimize cogging torque. Using only 6 644 finite‑element evaluations, the method reduced cogging torque by 94.75 % compared with the nominal design. In contrast, GP‑based RDO required roughly double the evaluations and yielded a smaller torque reduction; MC‑based RDO was computationally prohibitive.

Contributions

1. First integration of BNN and PDD for RDO, delivering both scalability to high dimensions and analytical moment evaluation.
2. An active‑learning scheme that leverages BNN predictive variance to concentrate expensive simulations where they are most needed.
3. A dynamic MPSS sub‑region strategy combined with global search to handle severe nonlinearity and avoid local‑optimum traps.

Limitations & Future Work
The quality of BNN uncertainty estimates depends on the variational approximation; more accurate inference (e.g., Hamiltonian Monte‑Carlo) could be explored. Selection of PDD truncation orders (interaction level S and polynomial degree m) remains heuristic; adaptive order selection would improve robustness. Finally, systematic studies on the impact of active‑learning thresholds and sub‑region resizing policies on convergence speed and solution quality are suggested.

Overall, the paper presents a compelling, computationally efficient RDO methodology that bridges probabilistic deep learning with classical polynomial decomposition, and validates its superiority on both synthetic and real‑world engineering problems.