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

Introduction

Designing high‑dimensional physical systems—such as laser pulses for inertial confinement fusion or data‑center workload schedules—often requires navigating expensive simulations and intricate constraints. In offline model‑based optimization (MBO), practitioners must improve designs using only a static dataset, without the luxury of querying the true objective during inference. Traditional offline MBO approaches either play it safe (conservatism), over‑exploit surrogate models, or act as static samplers that cannot be refined against precise physics‑based penalties. The recent Differentiable Diffusion Conditioning (D2C) framework offers a fresh perspective: it leverages the differentiability of the reverse denoising chain in conditional diffusion models to perform end‑to‑end gradient‑based optimization, even when the diffusion model itself remains frozen.

How D2C Works

At its core, D2C treats a pre‑trained conditional diffusion model as a proposer that generates candidate designs from a latent conditioning vector. Crucially, the reverse diffusion process—normally a deterministic sampling routine—can be differentiated with respect to this conditioning signal. By back‑propagating gradients from a frozen surrogate objective (trained on the offline dataset) and any additional physics‑informed penalty terms, D2C directly updates the conditioning variables to improve the generated designs.

The workflow consists of three stages:

1. Training the Conditional Diffusion Model – A standard diffusion model is trained on the available design data, learning to map conditioning variables (e.g., target performance metrics) to realistic design samples.
2. Freezing the Model – Once trained, the diffusion network’s weights are fixed, preserving its generative capabilities while allowing gradients to flow through its sampling steps.
3. Test‑Time Optimization – Starting from an initial conditioning vector, D2C iteratively computes the gradient of a composite loss (surrogate prediction + physics penalties) with respect to the conditioning. The conditioning is updated via gradient descent, and the diffusion sampler produces a refined design at each iteration.

Because the diffusion sampler is deterministic, each gradient step yields a concrete design rather than a stochastic sample, enabling stable convergence toward physically feasible solutions.

Applications and Results

The authors evaluated D2C on three distinct domains:

• Laser Pulse Optimization for Inertial Confinement Fusion – Here, the objective is to maximize energy delivery while respecting constraints on pulse shape and timing. D2C’s physics‑aware gradients allowed it to discover pulse configurations that outperformed baseline surrogate‑only methods.
• Sustainable Data‑Center Workload Scheduling – The task involves assigning workloads to servers to minimize energy consumption under latency constraints. By incorporating a penalty for server overheating, D2C produced schedules that reduced power usage without violating service‑level agreements.
• Design‑Bench Benchmark Suite – Across four synthetic design problems, D2C achieved the highest average rank, demonstrating its versatility beyond domain‑specific physics.

Across these experiments, D2C consistently delivered stronger results than competing offline MBO techniques, particularly on the two real‑world tasks where physics‑informed penalties play a critical role.

Practical Takeaways

• Leverage Existing Diffusion Models: D2C does not require retraining the generative model; any conditional diffusion model trained on the design dataset can be repurposed for optimization.
• Integrate Physics Directly: By formulating physical constraints as differentiable penalty terms, designers can steer the optimization toward feasible, real‑world solutions.
• Gradient‑Based Refinement: Unlike sampling‑only approaches, D2C’s gradient updates provide a clear path for iterative improvement, making it suitable for high‑stakes engineering problems where each design iteration must be justified.
• Modular Proposer‑Evaluator Architecture: The separation of a frozen proposer (diffusion model) and an evaluator (surrogate + physics penalties) simplifies experimentation—different evaluators can be swapped in without altering the generative backbone.

Conclusion

Differentiable Diffusion Conditioning bridges the gap between powerful generative diffusion models and the rigorous demands of physics‑informed offline optimization. By turning deterministic sampling into a differentiable conduit for gradient information, D2C enables designers to refine high‑dimensional systems against surrogate predictions and explicit physical constraints, all without needing test‑time access to the true objective. Its strong empirical performance across both real‑world and benchmark tasks suggests a promising new direction for model‑based optimization in scientific and engineering domains.