Surrogate Models for Stochastic Al-Cu Phase-Field Solidification

AI summary

Overview

Research area: Machine learning surrogates for computational materials science, specifically accelerated phase-field simulation of alloy solidification.

Technical level: Advanced. The paper assumes familiarity with partial differential equations, phase-field modeling, convolutional neural networks, and diffusion/denoising-bridge generative models.

Scope (one sentence): The paper introduces a variance-preserving denoising diffusion bridge model (VP-DDBM) that learns to predict the next state of a stochastic two-dimensional Al–Cu phase-field solidification simulation and compares it against two deterministic deep-learning baselines over long autoregressive rollouts.

What This Paper Is About

Phase-field simulations capture how solid–liquid interfaces and copper concentration evolve during rapid alloy solidification, but they are expensive, which limits long runs, large parameter sweeps, and ensembles of simulations. Because the simulator injects Gaussian random perturbations into the phase update, several future microstructures are plausible from the same starting state, so a purely deterministic surrogate cannot represent the full range of outcomes. The authors ask whether a probabilistic (stochastic) diffusion-bridge surrogate can predict the next stored simulation state, be applied recursively, and stay physically faithful over long horizons while running far faster than the original solver.

Key Contributions

1. A state-conditioned surrogate formulation for stochastic Al–Cu solidification: given the current phase field, Cu concentration field, and a prescribed temperature field, the model predicts the next state (750 phase-field Euler steps later) and is applied autoregressively. Temperature is treated as an input and is never predicted.

2. The introduction of VP-DDBM, a variance-preserving denoising diffusion bridge model that learns a conditional distribution over next states by training on stochastic bridges between known consecutive states, sampled as z_s = a_s x_k + b_s x_{k+1} + c_s ε, with the schedule satisfying z_1 = x_k and z_0 = x_{k+1}.

3. A shared-baseline comparison in which a deterministic U-Net, an SI-ConvNeXt, and VP-DDBM all use the same state variables, temperature information, trajectory split, 750-step transition interval, and autoregressive evaluation protocol, with evaluation on thermal-gradient families excluded from training.

4. A dual-pathway temperature conditioning design (FiLM modulation from the spatial mean and standard deviation of temperature, plus a five-resolution convolutional encoder injecting the full temperature field through zero-initialized 1×1 residual projections, analogous to ControlNet but trained jointly rather than frozen) combined with wall-clock benchmarking against the phase-field solver on stated hardware.

Main Findings

• Lower long-horizon error at the lower held-out gradient: At G = 1.6 × 10⁶ K m⁻¹, VP-DDBM produced lower errors than both deterministic baselines across all four tracked quantities — domain-mean phase, domain-mean Cu concentration, interface length, and leading-tip position — over 30,000 phase-field Euler steps, averaged across 15 held-out rollouts.

• Partial error advantage at the upper held-out gradient: At G = 2.2 × 10⁶ K m⁻¹, the same ordering held for domain-mean phase, domain-mean Cu concentration, and leading-tip position, but not for interface length.

• Morphology retention versus smoothing: VP-DDBM retained fine dendritic structure over long rollouts, whereas the deterministic U-Net progressively smoothed the morphology.

• Spatial fidelity in a long rollout: In the 65-transition rollout spanning simulator steps 244,500 to 293,250 (48,750 Euler steps), VP-DDBM reproduced similar front advancement and Cu redistribution while keeping a densely branched morphology.

• Speedup: For an equivalent 750-step advance, the median phase-field runtime was 46.29 s, versus 0.111 s for the deterministic U-Net and 0.634 s for VP-DDBM using five DBIM network evaluations — approximately 417× and 73× wall-clock speedups under the reported hardware configurations.

• Large parameter-count asymmetry: The deterministic U-Net used 214,154,338 trainable parameters, VP-DDBM 229,859,778, and SI-ConvNeXt only 1,265,459.

• Not reported numerically: The paper's figures plot relative errors over the rollouts, but the main text does not state explicit numeric error values for the four quantities; those are shown only in the plotted curves.

Methodology in Plain English

The authors first generated data by running a stochastic phase-field simulator for dilute Al–Cu on a uniform 512 × 768 grid: 70 independently generated trajectories at a pulling speed of 0.2 m s⁻¹, spanning 14 thermal gradients from 1.1 × 10⁶ to 2.4 × 10⁶ K m⁻¹ in increments of 0.1 × 10⁶ K m⁻¹, with five trajectories per gradient. This yielded 26,940 stored states and 26,870 adjacent transitions, each spanning 750 Euler steps.

Crucially, they split the data by whole thermal-gradient families, not by individual frames, so test conditions are genuinely unseen. Gradients 1.3 and 1.9 (in units of 10⁶ K m⁻¹) went to validation and 1.6 and 2.2 to testing, leaving ten gradients for training — producing 50, 10, and 10 trajectories with 19,275, 3,990, and 3,605 transitions respectively. Models were chosen by best validation loss.

Three models learn the same one-step mapping. The deterministic U-Net is a residual encoder–decoder with skip connections. SI-ConvNeXt uses resolution-preserving ConvNeXt blocks with no encoder–decoder bottleneck and takes phase, concentration, and temperature as three stacked channels. VP-DDBM instead learns a distribution: during training it creates noisy intermediate blends of the known current and next states and trains the network to recover the clean next state given the current state, the temperature field, and the bridge time. At inference, where the next state is unknown, the model walks the learned bridge from the current state toward a sampled next state using Diffusion Bridge Implicit Models, with a Gaussian perturbation applied near the start and deterministic updates thereafter.

A shared U-Net backbone design feeds temperature in two ways: global statistics through FiLM feature modulation, and the full temperature field through a five-resolution encoder injected at matching resolutions. Convolutions use periodic padding along y and nonperiodic padding along x to match simulator boundary conditions.

Evaluation rolls the models forward recursively — each prediction becomes the next input while the prescribed temperature field advances — and compares domain-mean phase, domain-mean Cu concentration, interface length (the one-dimensional Hausdorff measure of the zero contour of the phase field), and leading-tip position (the maximum x on that contour). Because stochastic forcing can shift individual sidebranches, the authors emphasize physically meaningful bulk and morphology quantities rather than pixel-wise agreement. Timings compared the solver on 12 CPU nodes of CSC's Mahti supercomputer with 1,536 MPI ranks against surrogate inference at batch size one on a single AMD MI250X Graphics Compute Die of LUMI.

Why This Matters

Impact on research: The work shows that a generative diffusion-bridge surrogate can preserve the stochastic character of a phase-field simulation rather than collapsing to a single smoothed mean trajectory, which matters for any PDE system with intrinsic noise. It also demonstrates a rigorous evaluation protocol — held-out thermal-gradient families, many rollouts, physically interpretable metrics — that is more informative than one-step pixel error.

Real-world applications:

• Screening process conditions for metal additive manufacturing, where thermal gradients and pulling speeds vary widely and full simulations are costly.
• Generating ensembles of microstructure predictions for statistical studies of dendritic morphology and sidebranching.
• Providing initial or intermediate microstructures for downstream property models that need many varied samples rather than one deterministic output.
• Reducing the cost of long solidification trajectories that would otherwise be impractical to run to completion.

Industry relevance: A roughly 73× reduction in wall-clock time for the stochastic model, and roughly 417× for the deterministic one, changes what is feasible in a design or qualification loop. The explicit hardware accounting (a supercomputer allocation versus a single GPU GCD) makes the trade-off concrete for organizations deciding whether surrogate deployment is worth the investment.

Future Directions

• Extending evaluation to a broader range of thermal histories and processing conditions relevant to additive manufacturing, as the authors propose.
• Understanding and improving the interface-length prediction at the upper held-out gradient, where VP-DDBM did not beat the deterministic baselines.
• Quantifying the quality of the stochastic ensemble itself — for example, whether sampled realizations cover the true distribution of sidebranch morphologies, which the paper treats qualitatively rather than numerically.
• Assessing whether the approach generalizes beyond two dimensions and beyond the specific Al–Cu system and parameter set used here.

Target Audience

Researchers and practitioners working at the intersection of scientific machine learning and computational materials science, particularly those using phase-field methods for solidification or additive-manufacturing modeling. It also suits machine-learning researchers interested in diffusion-bridge and other generative surrogates for spatiotemporal PDE systems, and computational scientists evaluating whether learned surrogates can replace or supplement expensive solvers in production workflows. Readers without a background in PDE simulation or diffusion models will find the methodology sections demanding.

Authors’ abstract

Phase-field models resolve microstructural evolution during rapid alloy solidification, but their computational cost limits long trajectories, stochastic ensembles, and studies across process conditions. We present a variance-preserving denoising diffusion bridge model (VP-DDBM) for state-conditioned prediction of two-dimensional Al-Cu solidification. Given the current phase, concentration, and prescribed temperature fields, the model predicts the subsequent state and is applied autoregressively. Across two held-out thermal gradients, VP-DDBM produces lower long-horizon errors than deterministic U-Net and SI-ConvNeXt baselines in domain-mean phase, domain-mean Cu concentration, and leading-tip position. It also retains fine dendritic structure over long rollouts, whereas the deterministic U-Net progressively smooths the morphology. VP-DDBM predicts the phase and concentration evolution approximately $73\times$ faster than the phase-field solver, showing potential for faster process-condition screening and ensemble-based microstructure studies.