Few-Shot Test-Time Optimization Without Retraining for Semiconductor Recipe Generation and Beyond

Shangding Gu, Donghao Ying, Ming Jin, Yu Joe Lu, Jun Wang, Javad Lavaei, Costas J. SpanosarXiv 2025 · DOI · arXiv

Abstract

In semiconductor manufacturing, achieving precise target specifications for processes like plasma etching is a resource-intensive task. Traditional approaches often involve slow manual tuning or computationally expensive model retraining. This paper presents Model Feedback Learning (MFL), a novel framework for few-shot test-time optimization. MFL optimizes inputs for pre-trained models or hardware systems without requiring any parameter updates to the underlying model. By iteratively refining a lightweight "reverse model" through a two-loop process, MFL generates complex manufacturing recipes in as few as five iterations, significantly outperforming existing optimization methods and human experts.

Introduction

The high-stakes environment of semiconductor fabrication demands extreme precision. When deploying equipment like plasma etching machines, engineers must determine the optimal "recipe"—a specific combination of gas flows, pressures, and temperatures—to achieve desired physical outcomes.

Semiconductor manufacturing process flow from recipe to target profiles.

As shown in the process flow above, the manufacturing sequence involves mapping a multi-parameter recipe to specific output metrics such as etch depth and critical dimensions (CD). Traditionally, finding the right recipe for a new target specification involves extensive trial-and-error or retraining complex AI models, both of which are costly and time-consuming.

The Deployment Gap and Inverse Problems

A significant hurdle in industrial AI is the "deployment gap." While a "forward model" or emulator can predict machine behavior given specific inputs, the practical challenge is the inverse: determining the inputs required for a specific target output. Existing solutions face several drawbacks:

  • Retraining: Computationally expensive and often impractical for deployed systems.
  • Bayesian Optimization (BO): Effective but slow, requiring numerous costly physical iterations.
  • Human Expertise: Senior engineers may require dozens of iterations to "dial in" a new recipe.

Methodology: Model Feedback Learning (MFL)

The MFL framework addresses these challenges by focusing on input optimization rather than model parameter optimization. The core component is a lightweight Reverse Model ($R$), a neural network (approx. 7 kB) designed to map desired targets back to optimal inputs.

The Model Feedback Learning (MFL) framework featuring Loop A (pre-training) and Loop B (test-time fine-tuning).

As illustrated in the framework diagram above, MFL operates through a two-loop training process:

  1. Loop A (Pre-training): The reverse model is trained against a digital emulator. This allows the system to learn the general physics of the process without consuming real-world resources.
  2. Loop B (Test-time Fine-tuning): Upon deployment, the reverse model is fine-tuned through direct interaction with the actual machine. The pre-training in Loop A ensures that Loop B requires only a few "shots" (iterations) to reach the target.

To ensure stability in sensitive process areas, MFL incorporates Conservative Learning. This mechanism monitors model sensitivity using Jacobian matrices; if the output is highly sensitive to input changes, the algorithm automatically reduces the learning rate to prevent unstable or damaging machine settings.

Experimental Validation

The researchers validated MFL across three industrial domains using high-fidelity simulations that incorporate real-world noise and variability:

  1. Semiconductor Plasma Etching: 11 input parameters and 6 target output metrics.
  2. Chemical Vapor Deposition (CVD): A process for producing high-quality solid materials.
  3. Wire Bonding: An electronics packaging process.

Key Results and Insights

MFL demonstrated superior performance in terms of speed, accuracy, and robustness:

  • Efficiency: Successful plasma etching recipes were generated in just 5 iterations, compared to approximately 84 for human experts and 20+ for Bayesian Optimization.
  • Precision: MFL consistently met strict target ranges for all metrics, including etch depth and critical dimensions, where baseline methods often failed.
  • Robustness: The system successfully adapted to simulated process shifts and "attack noise," maintaining stability where traditional supervised learning methods struggled.
  • Versatility: The framework's effectiveness extended beyond semiconductors, succeeding in CVD (5 iterations) and wire bonding (9 iterations).

Conclusion and Future Work

Model Feedback Learning offers a scalable, model-agnostic solution for the inverse problems prevalent in industrial manufacturing. By optimizing inputs through a lightweight reverse model, it bridges the gap between theoretical AI and practical engineering. Future research will focus on validating MFL on physical fabrication lines and extending the framework to reinforcement learning or fully offline scenarios.

Summary updated

Abstract

We introduce Model Feedback Learning (MFL), a novel test-time optimization framework for optimizing inputs to pre-trained AI models or deployed hardware systems without requiring any retraining of the models or modifications to the hardware. In contrast to existing methods that rely on adjusting model parameters, MFL leverages a lightweight reverse model to iteratively search for optimal inputs, enabling efficient adaptation to new objectives under deployment constraints. This framework is particularly advantageous in real-world settings, such as semiconductor manufacturing recipe generation, where modifying deployed systems is often infeasible or cost-prohibitive. We validate MFL on semiconductor plasma etching tasks, where it achieves target recipe generation in just five iterations, significantly outperforming both Bayesian optimization and human experts. Beyond semiconductor applications, MFL also demonstrates strong performance in chemical processes (e.g., chemical vapor deposition) and electronic systems (e.g., wire bonding), highlighting its broad applicability. Additionally, MFL incorporates stability-aware optimization, enhancing robustness to process variations and surpassing conventional supervised learning and random search methods in high-dimensional control settings. By enabling few-shot adaptation, MFL provides a scalable and efficient paradigm for deploying intelligent control in real-world environments.
BibTeX
@misc{gu2025few,
  author = {Shangding Gu and Donghao Ying and Ming Jin and Yu Joe Lu and Jun Wang and Javad Lavaei and Costas J. Spanos},
  title = {{Few-Shot Test-Time Optimization Without Retraining for Semiconductor Recipe Generation and Beyond}},
  year = {2025},
  eprint = {2505.16060},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.16060},
}
Figure