When a metallurgist says "anneal this alloy at 900 K," the sentence does not say how the microstructure will move. Whether the two metals stay mixed, separate into a fine sponge-like pattern, or coarsen into droplets depends on the shape of the alloy's free-energy curve at that composition and temperature. CALPHAD (CALculation of PHAse Diagrams) databases are fitted thermodynamic models that compute exactly this curve. Video generation has met the same problem in a different setting. Text underspecifies motion, and the field's working fix has been to insert an intermediate representation, such as optical flow, camera rays or layout boxes, between the prompt and the pixels.
This note proposes the materials version of that fix. A diffusion model (a generator that learns to turn random noise into samples by gradual denoising) forecasts microstructure evolution. At every step it is conditioned on per-pixel fields of chemical potential and free-energy curvature. A CALPHAD Gibbs-energy model computes these fields from the current frame's composition map. The conditioning is not a text prompt, and it is not a list of scalar process settings.
The claim is testable. The model should respond correctly when the thermodynamics change regime, for example when an alloy crosses the spinodal (the boundary inside which a uniform mixture becomes unstable and separates on its own) or when the free energy is perturbed by a few tens of meV/atom. Models conditioned on text or on scalar parameters should interpolate smoothly through such boundaries, which is the wrong behavior.
Core Claim
Direction 2 establishes that symbolic conditioning cannot pin down continuous dynamics. Overcoming the Limits of Textual Motion Descriptions and AnimateDiff show that sparse captions fail to specify motion. Motion-I2V responds by predicting dense optical flow before rendering. CameraCtrl responds by injecting camera geometry as Plücker ray coordinates.
These intermediate representations assume objects that keep their identity while they move. Phase separation breaks that assumption, because nothing is "moving" in the optical-flow sense: composition diffuses down chemical-potential gradients, and regions split and merge.
Our claim is that, for this class of dynamics, the correct analogue of a flow field is the field the governing equation itself consumes. That field is the local chemical potential $\mu(\mathbf{r}) = \partial G/\partial c$ evaluated at the current composition $c(\mathbf{r})$, together with the curvature $\partial^2 G/\partial c^2$. Here $G(c,T)$ is the CALPHAD molar Gibbs energy of the solution phase, and the curvature's sign marks whether a region is unstable.
Conditioning on these fields gives the generator the information that the description "anneal at 900 K" leaves out, in a form that varies across space and is recomputed as the microstructure changes. What would be new is the demonstration that this conditioning, rather than greater model capacity or a diffusion architecture as such, is what lets a learned surrogate get regime boundaries and out-of-distribution free energies right.
Why This Is a Real Gap and Not a Relabeling
Two papers motivate the gap. MicroEvoEval benchmarks 14 spatiotemporal models on phase-field tasks, including Cahn-Hilliard spinodal decomposition. Its models regress image sequences from image sequences with no thermodynamic input. They collapse at long horizons (PredRNN's SSIM falls to 0.263 on long-term grain growth), and the strongest model, SimVP.v2, degrades when physical parameters shift on the Dumbbell out-of-distribution set.
Accelerating CALPHAD-based Phase Diagram Predictions … Universal Machine Learning Potentials shows that phase-diagram topology changes qualitatively under energy errors as small as 50 meV/atom. A useful surrogate therefore has to be sharply sensitive to $G$ near regime boundaries and insensitive elsewhere. Nothing in MicroEvoEval tests that property.
Two other papers are adjacent but do different things. GrainPaint applies diffusion models to microstructure, but its outputs are static 3D grain fields with no thermodynamic input. CALPHAD-coupled phase-field work such as SiC growth uses real databases but needed 5,000 simulations for a sensitivity study, which is the cost a surrogate is meant to amortize.
A specific trap from video generation carries over in a sharper form. MotiF describes conditional image leakage: models conditioned on a first frame learn to copy it and ignore motion instructions. In microstructures, staying static is sometimes the correct answer, as for a metastable mixture outside the spinodal. Leakage and correct stasis are therefore indistinguishable unless the evaluation separates thermodynamic regimes. Our protocol is built around that separation.
Research Program
Data. We build a CALPHAD-parameterized Cahn-Hilliard dataset. $G$ uses a Redlich-Kister subregular solution form, the standard CALPHAD expression for a binary solution phase. We vary the interaction parameters, the temperature and the mean composition so that trajectories span three regimes: unstable (inside the spinodal), metastable (between the spinodal and the miscibility-gap boundary), and stable. The numerical solver is the ground truth.
The simulation grid, horizons and metrics follow MicroEvoEval's spinodal task, so results can sit beside its published models. We also run the original MicroEvoEval spinodal and Dumbbell splits unchanged. A second, harder dataset uses a ternary $G$. There the per-step thermodynamic evaluation is a larger share of solver cost, and this is where a surrogate's value is real rather than nominal.
Model. The model is a chunked autoregressive video diffusion network. It generates $k$ future composition frames from $m$ past frames plus three conditioning maps computed from the last observed frame: $\mu(\mathbf{r})$, $\partial^2 G/\partial c^2(\mathbf{r})$, and the CALPHAD equilibrium tie-line compositions at $T$. The conditioning maps are injected into the temporal layers, in the way CameraCtrl injects ray coordinates. Between chunks the maps are recomputed from the generated state, so the thermodynamics follows the evolving microstructure.
During sampling, each denoising step projects the sample onto the mean composition $\bar c$ of the input, which enforces mass conservation as a hard constraint. Morphology is left to the learned prior. This loosely borrows the idea in Time-to-Move of strong and weak constraint schedules for different parts of the output.
Sensitivity probe. We perturb $G$ by controlled offsets $\delta$ in the range 5–100 meV/atom, chosen so that some perturbations move a composition across the spinodal and others do not. We then measure whether the surrogate's predicted regime (whether separation starts, and its characteristic wavelength) tracks the solver's. Ideally the response is a sharp step at the boundary and flat elsewhere. This probe is the materials analogue of asking whether a video model actually obeys its control signal.
Baselines
The naive reference is persistence, which repeats the last observed frame. It scores perfectly on stable and metastable cases and fails on unstable ones. Reporting accuracy per regime against persistence exposes any model that wins only by leaking its input. A second reference point is the numerical phase-field solver itself, reported as wall-clock cost per trajectory. Any accuracy result is meaningless without the speed ratio.
The strongest existing learned competitor is the MicroEvoEval family, specifically SimVP.v2 and VMamba. They were the benchmark's most accurate and fastest models; VMamba runs grain-growth inference in 0.006 s. For spinodal decomposition, no reported result was found on Lacuna for an exact RMSE or SSIM value. The only figures in the file are PredRNN's long-term grain-growth SSIM of 0.263 and qualitative statements that RMSE grows by one to two orders of magnitude at long horizons and that SimVP.v2 degrades out of distribution.
We retrain these models on our data with the same frames, horizons, GPU budget and hyperparameter-search budget as our model. They get scalar conditioning (interaction parameters, $T$, $\bar c$) concatenated as channels, which is the fair version of "just give it the parameters."
A text-conditioned video diffusion baseline fine-tunes an AnimateDiff-style motion module on the same trajectories. Its captions state the processing condition ("quench to 900 K, Pt-rich composition"). It represents the textual-description route that Direction 2 argues is insufficient.
The key ablation runs our diffusion model three ways: with scalar parameter conditioning instead of the CALPHAD fields, with the fields but without recomputation between chunks, and without mass projection. A second ablation adds the CALPHAD fields to SimVP.v2. It decides whether the gain comes from the conditioning or from the generative architecture.
Our model beats the competition if three things hold. It must match or exceed SimVP.v2 on in-distribution L-ETAP and RMSE. It must cut the out-of-distribution error increase on held-out interaction parameters and on Dumbbell by a clear margin. On the sensitivity probe it must classify the regime correctly in the near-boundary cases where scalar-conditioned models fail.
Why not just…?
Why not just run the CALPHAD-coupled phase-field solver? For a single binary trajectory, the solver is cheap and exact, and we do not claim to beat it there. The surrogate earns its place only in two settings. One is large sweeps, like the 5,000-run sensitivity study in SiC growth. The other is multicomponent systems, where the thermodynamic evaluations dominate cost. We report the speed–accuracy curve on both the binary and the ternary dataset. If the ternary speedup turns out modest, the honest conclusion is that the surrogate is not worth building for these systems, and the paper will say so.
Why not just give SimVP.v2 or VMamba the physical parameters? This is the real threat. Parameter conditioning might be enough in distribution, and a deterministic regressor is simpler than a diffusion model. The SimVP.v2-plus-CALPHAD-fields ablation settles it. If that model matches our diffusion model, the contribution reduces to the conditioning representation, which we would then report as the finding. Our specific prediction is that scalar conditioning fails on the sensitivity probe and on held-out interaction parameters. A scalar input can only be mapped smoothly to outputs, while regime boundaries are thresholds in the local curvature, and the curvature field exposes those thresholds directly.
Why not fine-tune a text-to-video foundation model with descriptive captions? Evidence from MotiF and TTOM is that first-frame-conditioned models leak static content and miss compositional instructions without explicit intermediate control; TTOM needed layout-guided test-time optimization for an 82.57% motion-category gain on Wan2.1. We test the text-conditioned route directly. We expect it to look acceptable on aggregate SSIM, because persistence-like outputs score well on stable regimes, and to fail on per-regime accuracy.
Why not use universal machine-learned interatomic potentials instead? These potentials estimate energies of atomic configurations. They do not describe mesoscale evolution, so they complement this work rather than replace it. We use them as a stress test: building $G$ for one system from ORB-quality versus M3GNet-quality energies (32 K versus >900 K error on the Au-Pt critical temperature in the MLIP study) checks whether the surrogate propagates upstream thermodynamic error faithfully, which is what a downstream user needs.
Decisive Checks
The work stands or falls on two results.
The first is the regime-resolved comparison against persistence and SimVP.v2-with-scalars on the MicroEvoEval spinodal and Dumbbell splits plus our held-out interaction parameters. It must show gains concentrated in the unstable and near-boundary cases, measured by L-ETAP, wavelength error and onset time, rather than gains that come from copying stable inputs.
The second is the sensitivity probe. The predicted regime should flip within the same $\delta$ window as the solver's, near the 50 meV/atom scale that the MLIP study identifies as decisive for topology, and stay flat where the solver stays flat. If CALPHAD-field conditioning passes both checks and scalar conditioning fails them, we have shown that thermodynamic fields, not descriptions, are the control signal these generators need.
References
- AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning
- Motion-I2V: Consistent and Controllable Image-to-Video Generation with Explicit Motion Modeling
- CameraCtrl: Enabling Camera Control for Video Diffusion Models
- MicroEvoEval: A Systematic Evaluation Framework for Image-Based Microstructure Evolution Prediction
- Accelerating CALPHAD-based Phase Diagram Predictions in Complex Alloys Using Universal Machine Learning Potentials: Opportunities and Challenges
- GrainPaint: A multi-scale diffusion-based generative model for microstructure reconstruction of large-scale objects
- Phase-field model of silicon carbide growth during isothermal condition
- MotiF: Making Text Count in Image Animation with Motion Focal Loss
- Time-to-Move: Training-Free Motion-Controlled Video Generation via Dual-Clock Denoising
- TTOM: Test-Time Optimization and Memorization for Compositional Video Generation