End-to-End Differentiable Diffusion Conditioning for Physics-Informed Optimization

Optimizing high-dimensional physical systems under expensive evaluations and complex constraints is a central challenge in science and engineering. In offline model-based optimization (MBO), methods must improve designs using only a fixed dataset and no test-time access to the true objective.

在昂贵的评估和复杂的约束条件下优化高维物理系统是科学和工程领域的一项核心挑战。在离线基于模型的优化（MBO）中，方法必须仅使用固定数据集且在测试时无法访问真实目标函数的情况下改进设计。

However, existing methods typically suffer from conservatism, unconstrained surrogate exploitation, or function as static inference-time samplers that cannot refine candidates against precise physical objectives. In this paper, we propose Differentiable Diffusion Conditioning (D2C) for offline MBO, which leverages the fact that deterministic sampling makes the reverse denoising chain differentiable with respect to the conditioning signal.

然而，现有方法通常存在保守性、无约束的代理利用问题，或者仅作为静态推理时间采样器运行，无法根据精确的物理目标对候选方案进行优化。在本文中，我们为离线 MBO 提出了可微扩散条件化（D2C），它利用了确定性采样使得反向去噪链相对于条件信号可微这一事实。

D2C turns a frozen conditional diffusion model into an end-to-end gradient-based optimizer by directly optimizing conditioning variables at test time via gradients from frozen surrogate objectives and physics-informed penalties backpropagated through the reverse chain. We implement this idea in a composed proposer-evaluator framework and evaluate it on laser pulse optimization for inertial confinement fusion, sustainable data-center workload scheduling, and four Design-Bench tasks.

D2C 通过在测试时直接优化条件变量，将冻结的条件扩散模型转变为端到端的基于梯度的优化器，其梯度来自冻结的代理目标和通过反向链反向传播的物理信息惩罚。我们将这一思想实现在一个组合的提议者-评估者框架中，并在惯性约束聚变激光脉冲优化、可持续数据中心工作负载调度以及四个 Design-Bench 任务上进行了评估。

Our method achieves the strongest results on the two real-world tasks and the best average rank across four Design-Bench tasks.

我们的方法在两个现实世界任务中取得了最强的结果，并在四个 Design-Bench 任务中获得了最佳平均排名。