WorksComputer Science
Surrogate Models for Stochastic Al-Cu Phase-Field Solidification
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 compare a deterministic U-Net with conditional flow matching (Flow), conditional diffusion (EDM), and a variance-preserving denoising diffusion bridge (VP-DDBM) for two-dimensional Al-Cu solidification. Given the current phase, concentration, and prescribed temperature fields, each model predicts the subsequent state and is applied autoregressively. At two thermal gradients outside the training range, we compare surrogate rollouts spanning phase-field Euler steps with five phase-field futures generated from a common restart at each gradient. With two network evaluations per transition, Flow achieves the lowest time-averaged fair continuous ranked probability score among the generative surrogates for mean phase, mean Cu concentration, interface length, and leading-tip position at both gradients. Flow also reproduces the cellular morphology and retains sharp solid–liquid interfaces. It generates a single future approximately faster than the phase-field solver on the reported hardware, showing potential for faster process-condition screening and ensemble-based microstructure studies.
Introduction
Phase-field (PF) models resolve diffuse solid-liquid interfaces and their coupling to solute transport. Formulations have been developed for both near-equilibrium alloy solidification [1, 2] and rapid solidification [3, 4], including conditions relevant to metal additive manufacturing [5]. However, resolving the phase-field equations at the required spatial and temporal scales makes long trajectories, stochastic ensembles and large parameter studies computationally expensive.
Machine-learning surrogates can reduce the cost of PDE-governed simulation, as demonstrated in fluid dynamics and weather forecasting [6, 7, 8, 9]. Convolutional, recurrent and operator-based surrogates have also been developed for phase-field evolution [10, 11, 12, 13, 14].
In the Al–Cu simulator considered here, Gaussian noise in the phase update influences interfacial instabilities and dendritic sidebranching [15]. Different noise realizations can therefore produce different microstructures from the same state under the same thermal forcing. Probabilistic autoregressive models have been applied to stochastic mesoscopic crystallization [16], but stochastic surrogate prediction has received limited attention for coupled phase-solute evolution under a moving temperature field.
We therefore formulate surrogate modeling of stochastic Al-Cu solidification as state-conditioned forward prediction. Given the current phase and Cu concentration fields and the prescribed temperature field, each surrogate predicts the state 750 phase-field Euler steps later and is applied autoregressively. We compare a deterministic U-Net with three generative surrogates: conditional flow matching (Flow) [17], conditional diffusion (EDM) [18] and a variance-preserving denoising diffusion bridge (VP-DDBM) [19]. All four use residual U-Net backbones of approximately matched capacity, differing mainly in the training objective and the sampler.
Because a stochastic simulator has many valid futures, a single reference trajectory cannot show whether a surrogate reproduces its variability. At two thermal gradients outside the training range, we therefore compare ensembles of surrogate rollouts with ensembles of PF futures generated from the same restart state using independent solver noise. The initial microstructure and the thermal forcing are thus fixed, and the PF futures differ only through the noise. Noise can displace individual cells, so the comparison uses physical observables and their spread across futures, not pixel-wise agreement, together with the measured inference cost. This comparison tests whether a generative surrogate can reproduce the distribution of PF futures at a fraction of the solver cost.
Phase-field model for rapid solidification
At the continuum level, the dilute-alloy phase-field model follows the formulations in Refs. [3, 20], which describe the co-evolution of a solid-liquid order parameter () and alloying element concentration field () under a prescribed temperature field . The order parameter equation is of the form
where is the interface normal, and are anisotropic interface width and attachment time, respectively, is an additional anisotropy term, is the driving force term, and is a Gaussian stochastic forcing term following Ref. [15].
The concentration follows a nonlinear diffusion equation according to
where is the solute diffusion coefficient in the liquid phase, is an interpolation function, is the chemical potential, and is a prefactor to an antitrapping current used to impose controlled rapid solidification kinetics.
The imposed temperature field is prescribed as a moving linear temperature profile,
where is the solidification direction coordinate, is an appropriate reference shift, is the imposed thermal gradient, and is the pulling speed. The PF simulator uses an explicit Euler time integration scheme with a constant time step size . Details of the phase-field implementation and model parameters for the dilute Al–Cu alloy are given in Ref. [21].
Surrogate models
Let denote the phase and Cu concentration fields at stored state , and let denote the prescribed temperature field. All surrogates model the one-step transition from to . Temperature is prescribed and is not predicted.
We compare a deterministic U-Net with three generative surrogates: conditional flow matching (Flow), conditional diffusion (EDM), and a variance-preserving denoising diffusion bridge (VP-DDBM). The deterministic U-Net predicts using a residual encoder–decoder architecture with skip connections [22]. It is trained by minimizing mean-squared error against the next stored state and returns a single prediction. The generative surrogates instead model a conditional distribution over possible next states.
Conditional flow matching (Flow) [17] directly learns a velocity field along a prescribed path from Gaussian noise to the conditional distribution of next states. Conditional diffusion (EDM) [18] instead learns to reconstruct the clean next state from noise-corrupted versions at different noise levels. Its denoiser output is then used to construct the velocity of a probability-flow ODE. In our implementation, both samplers start from Gaussian noise and use Euler integration. EDM returns the output of the final denoiser evaluation. VP-DDBM models the conditional distribution using the variance-preserving formulation of Denoising Diffusion Bridge Models [19]. During training, the known current and next states form the endpoints of a stochastic bridge. An intermediate bridge state is sampled as
where is the bridge time and , , and are determined by the variance-preserving schedule. The schedule satisfies and , with vanishing at both endpoints. Intermediate states are therefore noisy mixtures of the current and next states. The network is trained to recover the clean next state from , conditioned on , , and .
At inference, is unknown. We use Diffusion Bridge Implicit Models (DBIM) [23] to traverse the learned bridge from the known current-state endpoint toward a sampled next state. A first network call at estimates the next state. This estimate, the current state, and a Gaussian boot perturbation initialize the bridge just below . With , the remaining DBIM updates are deterministic given that perturbation.
All four surrogates use residual U-Net backbones with the same temperature-conditioning design. Temperature enters through two pathways. First, its spatial mean and standard deviation are embedded and used by FiLM to rescale and shift feature channels, providing global thermal conditioning [24]. Second, the complete temperature field is processed by a five-resolution convolutional encoder, and its features are injected at the corresponding U-Net resolutions through zero-initialized residual projections. These projections are analogous to the zero-convolution connections used by ControlNet [25]. Unlike ControlNet, however, our architecture does not freeze or duplicate a pretrained backbone; all backbone and conditioning parameters are trained jointly. The generative networks also receive the intermediate sampler state, and the global temperature representation is combined with a sampling-time or noise-level embedding before feature modulation. Spatial convolutions use periodic padding along and nonperiodic padding along , consistent with the simulator boundary conditions.
The number of network function evaluations (NFE) counts all network calls per transition, including the final EDM denoiser call and the initial DBIM call. The deterministic U-Net uses one evaluation. Appendix § 6.§ 1 summarizes model sizes, training hyperparameters, and sampler settings.
Training and evaluation
The training dataset contains 70 phase-field trajectories on a uniform grid. Simulations use a pulling speed of 0.2 m/s and 14 thermal gradients from to K/m in increments of K/m, with five independently seeded trajectories per gradient.
We train three networks per method on LUMI [26] using different seeds and the same optimizer, batch size, update count, and weight-averaging scheme. Checkpoints are selected using only the prediction error of the ensemble mean on validation states at K/m. For U-Net, this error is computed from its single prediction.
We evaluate autoregressive rollouts from one restart state at each test gradient, K/m. The phase-field solver generates five futures per restart using independent noise. Each generative network produces five futures, and each U-Net produces one. Each rollout spans 48,000 phase-field Euler steps, with 750 Euler steps per surrogate transition.
We compare mean phase, domain-mean Cu concentration, interface length [27], and leading-tip position [28], together with qualitative spatial comparisons. For each observable, we report the fair continuous ranked probability score (CRPS) [29], averaged over 64 rollout horizons. We also compare the ensemble spread of interface length and leading-tip position. The score has the unit of the observable and is lower for a better ensemble. For U-Net, we report absolute error, equivalent to the CRPS of a single deterministic prediction. At the final horizon, we also compare mean cell spacing and its within-field standard deviation at the higher test gradient.
For the wall-clock comparison, the phase-field solver was timed on two CPU nodes of CSC’s Roihu supercomputer [30], each with two 192-core AMD EPYC 9965 processors. The solver used 512 MPI ranks, with 256 per node and one physical core per rank. Surrogate inference was measured at batch size one on a single NVIDIA GH200 GPU with 96 GiB of GPU memory. Solver times are medians across the five phase-field futures and include frame output. For each network, inference time is the median of three timed runs after one warm-up. Reported method timings are the median over the three networks and exclude input and output. All times are reported per surrogate transition.
Results
Figure 1 compares the phase-field reference with one autoregressive rollout from each surrogate at the higher test gradient. After 23,250 Euler steps, the reference front has developed a cellular array. Flow and VP-DDBM reproduce a similar cellular morphology. The U-Net front advances further than the reference, while EDM produces a strongly curved front that the reference does not show.

Figure 1. Phase-field reference and autoregressive surrogate rollouts at . Rows show the common initial state and the phase field after 23,250 and 48,000 phase-field Euler steps. All panels show the same crop along the growth direction and the middle half of the transverse domain. The generative models use two network evaluations per transition and the U-Net uses one.
Figure S4 shows the phase and Cu concentration fields at both test gradients. In the rollouts shown, Flow retains sharp solid–liquid interfaces comparable to the phase-field reference at both displayed times. EDM produces broader interfaces, while VP-DDBM produces broader interfaces at the intermediate horizon but recovers sharp interfaces by the final horizon. Figure S2 and Figure S3 compare
Figure S2. Autoregressive rollout scores and ensemble spread at the higher test gradient. Rollouts start at simulator step 131,250 at K/m and span 64 surrogate transitions, corresponding to 48,000 phase-field Euler steps. Panels (a)–(d) show scores for mean phase, mean Cu concentration, interface length, and leading-tip position, respectively. Scores are fair CRPS for the generative models and absolute error for the U-Net. At each scored horizon, scores are averaged first over the five phase-field futures and then over the three trained networks. Panels (e) and (f) show the standard deviation of interface length and leading-tip position, respectively, across the futures of each generative network, averaged over the three networks. Dashed black curves show the phase-field leave-one-out scores; solid black curves show the standard deviation across the five phase-field futures. The generative models use two network evaluations per transition and the U-Net uses one.
Figure S3. Autoregressive rollout scores and ensemble spread at the lower test gradient. Rollouts start at simulator step 401,250 at K/m and span 64 surrogate transitions, corresponding to 48,000 phase-field Euler steps. Panels (a)–(d) show scores for mean phase, mean Cu concentration, interface length, and leading-tip position, respectively. Panels (e) and (f) show the standard deviation of interface length and leading-tip position across futures. Score definitions, averaging, and black-curve conventions are the same as in Figure S2. The generative models use two network evaluations per transition and the U-Net uses one.
Figure S4. Representative held-out rollouts at both test gradients. The upper plot shows , with rollouts starting at simulator step 401,250. The lower plot shows , with rollouts starting at simulator step 131,250. Within each plot, the upper block shows the phase field and the lower block shows Cu concentration in wt. %. Columns show the phase-field reference, deterministic U-Net, Flow, EDM, and VP-DDBM. Rows within each block show the common initial state and the states after 23,250 and 48,000 phase-field Euler steps, corresponding to 31 and 64 surrogate transitions, respectively. Each generative model column shows one sampled realization using two network evaluations per transition.
the rollout scores and ensemble spread at the higher and lower test gradients, respectively. With two network evaluations per transition, Flow achieves the lowest fair CRPS among the generative surrogates for all four observables at both gradients, averaged over the scored horizons. Its scores are also below the absolute errors of the deterministic U-Net. At the higher gradient, Flow’s interface-length CRPS of 8.3 is close to the phase-field leave-one-out reference of 7.4 . Its scores remain above this reference for the other observables and at the lower gradient, so the results do not establish that Flow reproduces the full distribution of phase-field futures.
The agreement in ensemble spread is less consistent. At the lower gradient, Flow overestimates the spread of leading-tip position, whereas VP-DDBM is closer to the phase-field reference. For primary cell spacing at the higher gradient, Flow is the generative surrogate closest to the reference in both the mean and the within-field standard deviation at the final horizon (Figure S1).
Figure S1. Primary cell spacing at the higher test gradient. The left and right panels show the mean and standard deviation of cell spacings within each field, respectively, after 48,000 phase-field Euler steps at K/m. The black line and gray band show the mean and one standard deviation across five phase-field futures. Markers and bars show the mean and range across three trained networks, with five futures averaged per generative network and one prediction per U-Net. The generative models use two network evaluations per transition and the U-Net uses one.
For an equivalent advancement of 750 phase-field Euler steps, the median phase-field runtime is approximately 31.5 s, compared with 0.292–0.294 s for Flow across the two test gradients. On the reported hardware, this corresponds to a wall-clock speed-up of approximately . The other generative surrogates achieve similar speed-ups, while the deterministic U-Net achieves 211–215.
At two network evaluations per transition, Flow achieves the lowest averaged scores for all four physical observables at both test gradients. The representative rollouts also reproduce the reference cellular morphology and retain sharp solid–liquid interfaces (Figure S4). Together with the reduced rollout cost, these results make conditional flow matching a promising surrogate for ensemble studies of microstructural evolution during rapid Al–Cu solidification.
Supplementary materials
Model and training hyperparameters
All four surrogates were trained with AdamW, a learning rate of , a weight decay of , and an effective global batch size of 128. Training used exponential weight averaging and seeds 7, 42, and 1337. Model-specific settings are summarized in Table S1.
| Model | Parameters | NFE |
| Deterministic U-Net | 229,973,858 | 1 |
| Flow | 229,859,778 | 2 |
| EDM | 229,859,778 | 2 |
| VP-DDBM | 229,859,778 | 2 |
Table S1. Model-specific training and inference settings. NFE denotes the total number of network evaluations per transition in the reported comparison, including the final EDM denoiser call and the initial DBIM call.
Flow and EDM use Euler integration. VP-DDBM uses the DBIM sampler with an initial Gaussian perturbation and .
Each network was trained on 16 LUMI-G nodes.
Evaluation details
Let denote the grid points of the stored raster, and let denote the physical domain. The domain-mean phase and Cu concentration are defined as
where is the physical-domain area represented by grid point . The mean Cu concentration is reported in wt. %.
Following the phase-field convention, the solid-liquid interface is identified with the zero contour of the phase field [31]:
In two dimensions, the interface length is defined as
where denotes one-dimensional Hausdorff measure. Contour lengths are summed on the stored raster without smoothing or joining contours across the periodic boundary. Taking as the growth direction, the leading-tip position is defined as
Primary cells are detected as peaks of the solid-front profile along the periodic direction, using a minimum prominence of and a minimum separation of four raster rows. For detected cells, the mean spacing is
Figure S1 compares primary cell spacing at the higher test gradient. Figure S2 and Figure S3 report rollout scores and ensemble spread at the higher and lower test gradients, respectively. Figure S4 provides a spatial comparison of the phase and Cu concentration fields at both gradients.
References
- [1]Alain Karma. Phase-field formulation for quantitative modeling of alloy solidification. Physical Review Letters, 87(11):115701, 2001. doi: 10.1103/PhysRevLett.87.115701.DOI
- [2]Blas Echebarria, Roger Folch, Alain Karma, and Mathis Plapp. Quantitative phase-field model of alloy solidification. Physical Review E, 70(6):061604, 2004. doi: 10.1103/PhysRevE.70.061604.DOI
- [3]Tatu Pinomaa and Nikolas Provatas. Quantitative phase-field modeling of solute trapping and continuous growth kinetics in quasi-rapid solidification. Acta Materialia, 168:167–177, 2019. doi: 10.1016/j.actamat.2019.02.009.DOI
- [4]Kaihua Ji, Mingwang Zhong, and Alain Karma. Phase-field model of alloy solidification far from chemical equilibrium at the solid–liquid interface. Physical Review Research, 7(3):033128, 2025. doi: 10.1103/vyd6-nj4h.DOI
- [5]Neng Ren, Jun Li, Ruiyao Zhang, Chinnapat Panwisawas, Mingxu Xia, Hongbiao Dong, and Jianguo Li. Solute trapping and non-equilibrium microstructure during rapid solidification of additive manufacturing. Nature Communications, 14(1):7990, 2023. doi: 10.1038/s41467-023-43563-x.DOI
- [6]Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations, 2021. URL https://arxiv.org/abs/2010.08895.
- [7]Dmitrii Kochkov, Jamie A. Smith, Ayya Alieva, Qing Wang, Michael P. Brenner, and Stephan Hoyer. Machine learning–accelerated computational fluid dynamics. Proceedings of the National Academy of Sciences, 118(21):e2101784118, 2021. doi: 10.1073/pnas.2101784118.DOI
- [8]Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopadhyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, Pedram Hassanazadeh, Karthik Kashinath, and Anima Anandkumar. FourCastNet: A global data-driven high-resolution weather model using adaptive fourier neural operators, 2022. URL https://arxiv.org/abs/2202.11214.
- [9]Phillip Lippe, Bastiaan S. Veeling, Paris Perdikaris, Richard E. Turner, and Johannes Brandstetter. PDE-refiner: Achieving accurate long rollouts with neural PDE solvers. In Advances in Neural Information Processing Systems, volume 36, pages 67398–67433, 2023. URL https://proceedings.neurips.cc/paper_files/paper/2023/hash/d529b943af3dba734f8a7d49efcb6d09-Abstract-Conference.html.arxiv.org/abs/2308.05732
- [10]David Montes de Oca Zapiain, James A. Stewart, and Rémi Dingreville. Accelerating phase-field-based microstructure evolution predictions via surrogate models trained by machine learning methods. npj Computational Materials, 7(1):3, 2021. doi: 10.1038/s41524-020-00471-8.DOI
- [11]Christophe Bonneville, Nathan Bieberdorf, Arun Hegde, Mark Asta, Habib N. Najm, Laurent Capolungo, and Cosmin Safta. Accelerating phase field simulations through a hybrid adaptive Fourier neural operator with U-net backbone. npj Computational Materials, 11(1):14, 2025. doi: 10.1038/s41524-024-01488-z.DOI
- [12]Christophe Bonneville, Nathan Bieberdorf, Pieterjan Robbe, Mark Asta, Habib N. Najm, Laurent Capolungo, and Cosmin Safta. Extrapolating phase-field simulations in space and time with purely convolutional architectures, 2025. URL https://arxiv.org/abs/2509.20770. NeurIPS 2025 Workshop: ML for the Physical Sciences.
- [13]Kaihua Ji, Luning Sun, Shusen Liu, Fei Zhou, and Tae Wook Heo. Scalable autoregressive deep surrogates for complex microstructure dynamics. Machine Learning: Science and Technology, 7(4):045050, 2026. doi: 10.1088/2632-2153/ae94e4.DOI
- [14]Damien Pinto, Michael Greenwood, and Nikolas Provatas. LeapFrog: Getting the jump on multi-scale materials simulations using machine learning, 2024. URL https://arxiv.org/abs/2406.15326.
- [15]Alain Karma and Wouter-Jan Rappel. Phase-field model of dendritic sidebranching with thermal noise. Physical Review E, 60(4):3614–3625, 1999. doi: 10.1103/PhysRevE.60.3614.DOI
- [16]Pol Timmer, Koen Minartz, and Vlado Menkovski. Efficient probabilistic modeling of crystallization at mesoscopic scale. In Sašo Džeroski, Jurica Levatić, Gianvito Pio, and Nikola Simidjevski, editors, Discovery Science, volume 16090 of Lecture Notes in Computer Science, pages 3–16, Cham, 2025. Springer. doi: 10.1007/978-3-032-05461-6_1.DOI
- [17]Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. In International Conference on Learning Representations, 2023. URL https://arxiv.org/abs/2210.02747.
- [18]Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, volume 35, pages 26565–26577, 2022. URL https://arxiv.org/abs/2206.00364.
- [19]Linqi Zhou, Aaron Lou, Samar Khanna, and Stefano Ermon. Denoising diffusion bridge models. In International Conference on Learning Representations, 2024. URL https://arxiv.org/abs/2309.16948.
- [20]Tatu Pinomaa, Joseph M. McKeown, Jörg M. K. Wieczorek, Nikolas Provatas, Anssi Laukkanen, and Tomi Suhonen. Phase field modeling of rapid resolidification of Al–Cu thin films. Journal of Crystal Growth, 532:125418, 2020. doi: 10.1016/j.jcrysgro.2019.125418.DOI
- [21]Tatu Pinomaa, Adriana Eres-Castellanos, Pascal Thome, Joni Kaipainen, I-Ting Ho, Scott Hunter, Jeff Squier, Kamel Fezzaa, Abhishek Biswas, Sammy Tin, Nikolas Provatas, Amy J. Clarke, Joseph McKeown, and Thomas Voisin. Multi-scale in situ experiment/simulation integration to predict rapid solidification microstructures in additively manufactured AlCu, 2025. URL https://ssrn.com/abstract=5402093. SSRN preprint.
- [22]Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, volume 9351 of Lecture Notes in Computer Science, pages 234–241. Springer, 2015. doi: 10.1007/978-3-319-24574-4_28.DOI
- [23]Kaiwen Zheng, Guande He, Jianfei Chen, Fan Bao, and Jun Zhu. Diffusion bridge implicit models. In International Conference on Learning Representations, 2025. URL https://arxiv.org/abs/2405.15885.
- [24]Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. FiLM: Visual reasoning with a general conditioning layer. Proceedings of the AAAI Conference on Artificial Intelligence, 32(1):3942–3951, 2018. doi: 10.1609/aaai.v32i1.11671.DOI
- [25]Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3813–3824. IEEE, 2023. doi: 10.1109/ICCV51070.2023.00355.DOI
- [26]LUMI Consortium. GPU nodes – LUMI-G. https://docs.lumi-supercomputer.eu/hardware/lumig/, 2026. Accessed 2026-08-30.
- [27]Anh Tran, Dehao Liu, Hoang A. Tran, and Yan Wang. Quantifying uncertainty in the process–structure relationship for Al–Cu solidification. Modelling and Simulation in Materials Science and Engineering, 27(6):064005, 2019. doi: 10.1088/1361-651X/ab2690.
- [28]Andrea M. Jokisaari, Peter W. Voorhees, Jonathan E. Guyer, James A. Warren, and Olle G. Heinonen. Phase field benchmark problems for dendritic growth and linear elasticity. Computational Materials Science, 149:336–347, 2018. doi: 10.1016/j.commatsci.2018.03.015.
- [29]Christopher A. T. Ferro. Fair scores for ensemble forecasts. Quarterly Journal of the Royal Meteorological Society, 140(683):1917–1923, 2014. doi: 10.1002/qj.2270.DOI
- [30]CSC – IT Center for Science. Roihu supercomputer. https://docs.csc.fi/computing/systems-roihu/, 2026. Accessed 10 November 2026.
- [31]Ying Sun and Christoph Beckermann. Sharp interface tracking using the phase-field equation. Journal of Computational Physics, 220(2):626–653, 2007. doi: 10.1016/j.jcp.2006.05.025.



