Diffusion Tree Sampling for Inference Time Adaptation of Crystal Generative Models
Diffusion Tree Sampling lets crystal generative models be steered at inference time, enabling multi‑objective optimization without retraining.

Introduction
Diffusion‑based generative models have emerged as powerful priors for proposing plausible crystal structures. Yet, turning these priors into useful materials often requires balancing several competing properties—stability, rare space‑group occupancy, high bulk modulus, dielectric constant, bandgap, and more. Traditional approaches rely on fine‑tuning a model for each new objective or on differentiable guidance that may not apply to many real‑world tasks. The paper Diffusion Tree Sampling for Inference Time Adaptation of Crystal Generative Models introduces a different strategy: adapt the model during inference using a Monte‑Carlo tree search (MCTS) over the denoising trajectory, called Diffusion Tree Sampling (DTS). This technique preserves the original model while allowing flexible, multi‑objective steering.
How Diffusion Tree Sampling Works
DTS treats each step of the diffusion denoising process as a decision point in a search tree. At each node, the algorithm evaluates multiple candidate latent updates (the "actions") and expands the most promising branches based on a scoring function that reflects the desired material properties. By repeatedly sampling and back‑propagating rewards, DTS builds a tree that guides the diffusion trajectory toward regions of latent space likely to yield crystals meeting the target criteria. Crucially, the method can incorporate any existing property‑conditional generative model—such as a classifier‑free guidance model—as a proposal distribution without altering the underlying probability flow, ensuring that the final sample distribution remains faithful to the original diffusion prior.
Extending DTS to Multi‑Objective Sampling
A major contribution of the work is the extension of DTS to handle multiple objectives simultaneously. Instead of optimizing a single scalar reward, the search maintains a Pareto front of partial solutions. Each node stores a vector of property scores, and the tree expansion uses a multi‑objective selection rule (e.g., hypervolume improvement) to balance trade‑offs. The result is a single inference run that produces a diverse set of crystals spanning the entire Pareto frontier—for example, simultaneously maximizing dielectric constant and bandgap. This capability is especially valuable in materials discovery, where designers often need to explore trade‑offs rather than a single optimum.
Demonstrations on Crystal Generation
The authors showcase three practical scenarios:
- Stability Enhancement – Applying DTS to several baseline diffusion models improves the thermodynamic stability of generated crystals, reducing the fraction of high‑energy structures without retraining the models.
- Targeting Rare Space Groups and High Bulk Modulus – By feeding a conditional model fine‑tuned for classifier‑free guidance into DTS, the search can bias samples toward under‑represented space groups or toward crystals with unusually high bulk modulus, achieving higher hit rates than naïve sampling.
- Pareto Front of Dielectric Constant vs. Bandgap – Using the multi‑objective extension, DTS generates a continuum of candidates that jointly optimize these two electronic properties, providing a ready‑to‑use library for downstream experimental screening.
Across all tasks, DTS delivers measurable improvements while leaving the original diffusion models untouched, highlighting its practicality for AI‑for‑science pipelines.
Practical Takeaways
- Inference‑Only Adaptation – Researchers can retrofit existing diffusion models with DTS, avoiding costly fine‑tuning cycles.
- Modular Integration – Any property‑conditional model can serve as a guided proposal, making DTS compatible with a wide range of pretrained generators.
- Multi‑Objective Exploration – The Pareto‑aware search yields diverse candidate sets, supporting decision‑making in materials design.
- Scalability – Because DTS operates on the denoising trajectory, it scales with the same computational budget as standard diffusion sampling, adding only the overhead of tree management.
Conclusion
Diffusion Tree Sampling offers a versatile, inference‑time toolkit for steering crystal generative models toward complex, multi‑property goals. By marrying Monte‑Carlo tree search with diffusion denoising, it sidesteps the need for repeated model retraining and opens a path to systematic, Pareto‑driven materials discovery. Researchers looking to extract more value from existing diffusion models should consider integrating DTS into their generation pipelines.
Source
Read original sourceWhy we picked this
Core AI research on adapting diffusion generative models for scientific discovery.