Generative Inverse Design of Metamaterials with Functional Responses by Interpretable Learning

Problem

Framing

Inverse design for metamaterials with functional responses still relies on iterative search or one-to-one inverse maps, both of which break under non-unique targets and small datasets. The paper closes this gap with RIGID, which turns a random-forest forward model into an interpretable likelihood over designs and samples multiple valid solutions without retraining.

Currently Used Methods

Foundational

• "Tandem neural networks for inverse design of nanophotonic structures" — cascaded inverse and pretrained forward networks for inverse design.
  • Limitation in context: still favors one output per target, so diversity stays weak.
• @DenoisingDiffusionProbabilisticModels2020 — conditional generative diffusion for one-to-many design synthesis.
  • Limitation in context: higher data demand and lower interpretability on small metamaterial datasets.
• @kingmaVAE2013 — latent-variable generation adapted to conditional inverse design.
  • Limitation in context: solution likelihood is not explicit in original design space.
• @goodfellowGAN2014 — adversarial generation behind cGAN inverse-design baselines.
  • Limitation in context: unstable training and weak transparency for target satisfaction.
• "Interpretable machine learning for inverse design of metamaterials" — decision-tree rules expose feasible design regions.
  • Limitation in context: needs an extra surrogate tree and lacks calibrated region likelihoods.

Proposed Method

Architecture

RIGID trains a random forest on $(\mathbf{x}, s) \mapsto y$, where $\mathbf{x}$ are design variables, $s$ is frequency or wavelength, and $y \in \{0,1\}$ marks qualitative target satisfaction. Inversion prunes each tree by the target domain $\Omega$, maps surviving leaves to design-space regions, and averages leaf probabilities across trees into a forest likelihood.

Loss / Objective

The objective is to find designs whose qualitative response stays positive over the target domain.

For tree $m$, the target-satisfaction likelihood is

and the forest likelihood is

Sampling Rule / Algorithm

Generation samples designs in proportion to the estimated likelihood, using MCMC over the forest score.

Training Procedure

• Forward task: binary classification of qualitative functional responses.
• Inputs: design variables $\mathbf{x}$ and auxiliary variable $s$.
• Model: random forest with $M$ trees.
• Inversion: prune branches inconsistent with target interval(s) in $s$.
• Region score: leaf positive-class proportion.
• Generation: MCMC sampling from $L(\mathbf{x}\mid T)$.

Evaluation

Datasets

• Acoustic metamaterial bandgap design.
• Optical metasurface high-absorbance design.
• Synthetic SqExp benchmark.
• Synthetic SupSin benchmark.

Metrics

• Satisfaction rate.
• Average score.
• Selection rate.
• Forward-model test F1 score.

Headline results

• Acoustic forward model: test F1 $0.82$.
• Optical forward model: test F1 $0.83$.
• SqExp forward model: test F1 $0.85$.
• SupSin forward model: test F1 $0.86$.
• Real metamaterial studies: demonstrated with training sets in the $<250$ regime.

Table 1: Optical metasurface inverse-design results and diagnostics from Figure 5 panels C-G.

Ablations

• Sampling threshold $\tau$: higher $\tau$ lowers selection rate and raises satisfaction rate.
• Sampling threshold $\tau$: average score also rises with $\tau$.
• Forest size: more trees improve likelihood estimation quality.
• RIGID vs GA: RIGID covers broader feasible regions on multimodal targets.

Method Strengths and Weaknesses

Strengths

• No separate inverse network is trained; inversion comes from the forward forest.
• Explicit $L(\mathbf{x}\mid T)$ gives interpretable target-satisfaction scores.
• Works on small-data real cases with sub-250-sample training sets.
• Covers broader feasible regions than GA on multimodal problems.

Weaknesses

• Handles binary qualitative targets, not full quantitative response matching.
• Likelihood is piecewise constant over leaf-defined regions.
• Generation still relies on MCMC rather than direct ancestral sampling.
• Baselines emphasize GA and synthetic tests, not stronger modern conditional generators.

Suggestions from the authors

• Extend RIGID to quantitative functional-response targets.
• Study higher-dimensional design spaces.
• Use the synthetic problems as standard inverse-design benchmarks.
• Expand design spaces when likelihood indicates poor target coverage.

Links

Prior Papers

• @DenoisingDiffusionProbabilisticModels2020 — diffusion is a cited conditional generative baseline for one-to-many inverse design.
• @goodfellowGAN2014 — adversarial generation underlies cGAN inverse-design baselines discussed by the paper.
• @kingmaVAE2013 — variational latent modeling underlies cVAE inverse-design baselines.
• @wuGroupNormalization2018 — normalization layers appear in the deep generative baselines surveyed here.

Further Papers

No vault papers identified as further work yet.