RID-Noise: Towards Robust Inverse Design under Noisy Environments

Jia-Qi Yang, Ke-Bin Fan, Hao Ma, De-Chuan Zhan

Introduction

In engineering, a "perfect" design that fails under environmental noise or manufacturing tolerances is impractical. Traditionally, achieving robust design—creating products insensitive to such perturbations—required grueling, iterative "evaluate-and-feedback" loops. A new framework proposed by researchers at Nanjing University, titled "RID-Noise: Towards Robust Inverse Design under Noisy Environments," offers a more efficient path. By combining generative modeling with a novel "predictability" metric, RID-Noise learns to identify stable design parameters directly from noisy historical data, bypassing the need for iterative refinement.

The Challenge of Noisy Inverse Design

Inverse design involves finding design parameters ($x$) that produce a target performance ($y$). While deep learning has accelerated this process, most models assume a "clean" environment. In reality, data is often contaminated with noise. A specific design $x$ might yield target $y$ in a controlled setting, but minor environmental fluctuations could lead to catastrophic failure. Standard generative models, such as conditional Variational Autoencoders (cVAEs) or Generative Adversarial Networks (cGANs), tend to fit the entire dataset, including these unstable points. The core challenge is enabling an AI to distinguish between "safe" and "dangerous" regions of a noisy dataset without explicit labels.

Comparison of traditional Robust Design (top) and the proposed Robust Inverse Design (bottom).

Methodology: RID-Noise

The central hypothesis of RID-Noise is that robustness is synonymous with predictability. The framework consists of three main components:

  1. Measuring Robustness via Predictability: The system first trains a standard forward neural network ($x \to y$). If the network can accurately predict the outcome of a design parameter $x$, that parameter is likely robust. Conversely, high prediction error indicates sensitivity to noise.
  2. Sample-Wise Weighting: Each data point in the training set is assigned a weight ($w$) based on its prediction error. Points with high predictability receive higher weights, while unstable points are down-weighted.
  3. Weighted Likelihood Training: The generative core uses a conditional Invertible Neural Network (cINN). Unlike GANs, cINNs allow for exact likelihood calculation. By training the cINN with a Weighted Negative Log Likelihood (WNLL) loss, the model naturally prioritizes the stable regions of the design space.
  4. Conservatism Control: A hyperparameter $\tau$ acts as a "conservatism" dial, allowing engineers to control how aggressively the model ignores noisy data.

Experimental Evaluation

The researchers evaluated RID-Noise on both synthetic toy benchmarks and real-world engineering tasks, including robotic kinematics, ballistics, and meta-material design.

Key Insights from Toy Benchmarks

In the Radian task, the model must avoid a "dangerous" discontinuity where the output $y$ jumps abruptly. RID-Noise successfully learned to stay away from these boundaries, whereas baselines like cVAE and cGAN often selected designs on the edge of failure.

Performance comparison on the Radian task. RID-Noise achieves significantly lower MSE by avoiding unstable regions.

Real-World Performance

RID-Noise consistently outperformed state-of-the-art methods across all benchmarks:

  • Kinematics: In robotic arm joint angle determination with input noise, RID-Noise achieved an MSE of 0.0874, a nearly 50% reduction compared to the standard cINN (0.1732).
  • Radius Task: RID-Noise achieved an MSE of 0.0129, significantly lower than the ~0.0530 achieved by cVAE and cINN.
  • Statistical Significance: T-tests confirmed that RID-Noise’s improvements were statistically significant ($p < 0.05$) across almost every noisy scenario.

Conclusion

RID-Noise represents a significant step forward in making inverse design practical for real-world, noisy environments. By leveraging predictability as a proxy for robustness, the framework allows for the direct generation of stable designs from imperfect data. While the method depends on the accuracy of the initial forward model, its ability to provide a distribution of robust candidates offers engineers a powerful new tool for reliable product development.

Paper Citation: Jia-Qi Yang, Ke-Bin Fan, Hao Ma, De-Chuan Zhan. "RID-Noise: Towards Robust Inverse Design under Noisy Environments." Proceedings of the AAAI Conference on Artificial Intelligence, 2022. arXiv:2112.03912

Summary updated

Abstract

From an engineering perspective, a design should not only perform well in an ideal condition, but should also resist noises. Such a design methodology, namely robust design, has been widely implemented in the industry for product quality control. However, classic robust design requires a lot of evaluations for a single design target, while the results of these evaluations could not be reused for a new target. To achieve data-efficient robust design, we propose Robust Inverse Design under Noise (RID-Noise), which can utilize existing data to train a conditional invertible neural network. Specifically, we estimate the robustness of a design parameter by its predictability, measured by the prediction error of a forward neural network. We also define a sample-wise weight, which can be used in the maximum weighted likelihood estimation of an inverse model based on a conditional invertible neural network. With the visual results from experiments, we clearly justify how RID-Noise works by learning the distribution and robustness from data. Further experiments on several real-world benchmark tasks with noises confirm that our method is more effective than other state-of-the-art inverse design methods. Code and supplementary is publicly available at https://github.com/ThyrixYang/rid-noise-aaai22
BibTeX
@inproceedings{yang2022rid,
  author = {Jia-Qi Yang and Ke-Bin Fan and Hao Ma and De-Chuan Zhan},
  title = {{RID-Noise: Towards Robust Inverse Design under Noisy Environments}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i4.20390},
}
Figure