Data-efficient continuous conditional denoising diffusion model for microstructure generation

Jul 16, 2026

Traditional microstructure simulation methods like cellular automata and phase-field approaches can model evolution, yet often suffer from computational bottlenecks that hinder high-throughput and on-demand process optimization. A continuous conditional denoising diffusion model is proposed to generate microstructures conditioned on continuous processing parameters. Training on a compact process–microstructure dataset uses noise injection and progressive denoising to learn statistical structure. A vicinal-loss strategy improves data efficiency for sparsely sampled conditions, combined with classifier-free guidance and diffusion implicit sampling. Experiments on low-carbon steel conditioned on manganese composition reproduce key physical features including phase, grain morphology, size distribution, phase fraction, and interfacial area distribution.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages42

Reading Time1 h 46 min

Introduction

Deep learning for microstructure generation

Discriminative models are incompatible with stochastic microstructures

Generative modeling for stochastic microstructure generation

arXiv :2607 . 10429v 1 [ cs .CE] 11 Jul 2026 Data-efficient continuous conditional denoising diffusion model for microstructure generation Tarakram Ramgopala , Gowtham Nimmal Haribabua , Hussein Farahania,b , Cornelis Bosa,b , Siddhant Kumara a Department of Materials Science and Engineering, Delft University of Technology, 2628 CD Delft, The Netherlands b Tata Steel, Research & Development, P.O. Box 10000, IJmuiden, 1970 CA, Netherlands Abstract Traditional computational models, such as cellular automata and phase-field methods, are effective for simulating microstructural evolution but often face computational bottlenecks, limiting their application in high-throughput and on-demand process optimization. Generative machine learning approaches, such as denoising diffusion models, have emerged as powerful tools for surrogate modeling of process-structure maps, specifically producing representative microstructures conditioned on process parameters. However, they often require large amounts of data for training, particularly when process conditions are continuous rather than discrete categorical variables. To address this, we present a continuous conditional denoising diffusion model for generating microstructures conditioned on processing parameters. Trained on a compact dataset of process-microstructure pairs, this framework first adds noise to microstructure images and then trains a neural network to progressively remove the noise, learning the underlying statistical patterns of the microstructure. To address data inefficiencies associated with continuously valued process conditions, we propose a vicinal-loss training strategy that associates process conditions in data-sparse regions with nearby conditions in the dataset. Combined with classifier-free guidance and denoising diffusion implicit sampling, this approach enables data-efficient continuous conditional generation of microstructures compared to classical denoising diffusion models. The model successfully generates representative microstructures for low-carbon steel conditioned on manganese composition, matching key physical features such as phase and grain morphology, grain size distribution, phase fraction, and interfacial area distribution. More generally, this approach opens avenues for efficient process design and optimization of materials and their microstructures. 1. Introduction In metallic systems, processing parameters influence microstructural evolution, which in turn dictates material properties and performance. Conversely, achieving tailored properties requires control of the microstructure, which in turn necessitates process optimization. However, traditional trial-and-error methods for constructing process-structure maps are inefficient due to high experimental costs. While simulations based on, e.g., cellular automata (Yazdipour et al., 2008 ; Bos et al., 2010, 2011), phase field methods (Chen, 2002 ; Choi et al., 2024 ; Bhadeshia, 2014 ; Peivasteet al., 2022 ; Hu et al., 2022 ; Gao et al., 2023 ; Xue et al., 2022 ; de Oca Zapiain et al., 2021), and Potts-type Monte Carlo simulations (Hore et al., 2013 ; Tong et al., 2002) offer an efficient means to virtually explore process–structure relationships, their long computational runtimes often preclude real-time, process-conditioned microstructure generation and process optimization required in industrial metals production and processing. For example, a key circularity challenge is scrap steel recycling, where varying impurity levels affect material composition, requiring models that relate processing parameters to microstructural features for real-time control to achieve target properties. In such Email address: [Sid.Kumar@tudelft.nl](Sid.Kumar@tudelft.nl) (Siddhant Kumar) cases, long simulation times are impractical, and high-throughput, process-conditioned microstructure generation is required. Deep learning for microstructure generation: In recent years, machine learning (ML), spe

FAQs

Why are classical simulation methods often insufficient for process-conditioned microstructure generation?

They can accurately simulate microstructural evolution but may have long runtimes, preventing real-time or high-throughput process optimization and microstructure generation in industrial settings.

How does the proposed model generate microstructures conditioned on processing parameters?

It uses a continuous conditional denoising diffusion framework: noise is added to microstructure images, then a neural network progressively removes the noise while learning underlying microstructural statistical patterns conditioned on processing parameters.

What technique improves data efficiency when processing conditions are continuous and data-sparse?

A vicinal-loss training strategy associates conditions in sparse regions with nearby conditions present in the dataset, enabling efficient continuous conditional generation. It is combined with classifier-free guidance and denoising diffusion implicit sampling.

Jul 16, 2026

Upload by Theodore