Semiconductor TCAD Surrogates: FP-DDM Domain Decomposition and Etch Process Surrogates
NewOverview
Semiconductor process and device simulation is expensive enough that design exploration has to trade fidelity for iteration speed. NVIDIA's own technical account of AI-augmented TCAD notes that simulations can run for weeks, and that engineers need to evaluate tens of thousands of process cases generated from dozens of recipe combinations. That gap is where neural surrogates enter.
This entry anchors the discussion on two verifiable artifacts rather than on a general promise. The first is FP-DDM, a domain-decomposition framework proposed by a Samsung and UNIST team at DAC 2026, together with its open reference implementation inside the NVIDIA PhysicsNeMo repository. The second is the publicly documented method progression of the TCAD intelligence team at SK hynix, which built an etch-profile surrogate on top of PhysicsNeMo.
An earlier version of this entry described the TCAD examples as replacing repeated device-level numerical solves with data-driven models, and framed the target as thermal-electric co-design for 3D-IC and advanced packaging. Both descriptions were imprecise. The tcad directory currently contains a single sub-example, fp_ddm, which solves a steady two-dimensional thermal equation by overlapping Schwarz domain decomposition and compares against a two-dimensional plane-stress elasticity baseline. It does not solve drift-diffusion transport, does not model electrical behaviour, and operates in two dimensions. The industrial case that is actually on record concerns etching, which belongs to process TCAD rather than device TCAD.
Related material on this site covers the framework itself in tools/physicsnemos, the methodological limits of physics-constrained learning in guides/pinn-landscape-2026, and the general taxonomy of surrogates and reduced-order models in collections/surrogate-modeling-rom.
How It Works
FP-DDM expands to Foundation-model-based Physics-guided adaptation for the Domain Decomposition Method. The strategy partitions a large multi-physics problem into overlapping subdomains, uses a neural operator as the local solver on each subdomain, and enforces consistency across the artificial boundaries by explicitly optimising the values in the overlap regions through automatic differentiation.
The open reference implementation targets the steady two-dimensional heat equation with spatially varying conductivity, a volumetric heat source, and Dirichlet temperature boundary conditions. Each subdomain can be solved either by a bundled matrix-free finite-volume reference solver, configured through the fem key retained from the original scripts, or by a physics-informed PhysicsNeMo FNO.
The FNO is trained without solution labels. Its objective combines the thermal equation residual with the boundary residual, so changes in subdomain geometry, material distribution, and boundary conditions do not require fresh simulation data. Each decomposition iteration performs five steps: optionally adapting the FNO on the current subdomain batch with the same physics and boundary losses; solving every overlapping subdomain with the configured local solver; updating each artificial boundary from the neighbouring overlap; measuring interface consistency and assembling global diagnostics; and stopping on a tolerance, an iteration limit, or a patience limit.
Three interface handlers are available. The parallel handler optimises all artificial boundaries in a single batched step, the gradient handler updates patches sequentially, and the exchange handler performs a direct relaxed Dirichlet exchange. Overlap RMSE is the stopping metric, and the documentation states clearly that reaching the maximum iteration count is an iteration-budget limit rather than convergence.
The elasticity path is a numerical baseline. It uses manufactured solutions for a two-dimensional plane-stress bending problem with prescribed displacement on the entire outer boundary, compares a two-by-two overlapping decomposition against a monolithic solve using the same matrix-free operator, and prints relative displacement and stress errors. It trains no elasticity FNO and therefore does not constitute a neural FP-DDM result.
Scope and Assumptions
The repository describes this example as a research implementation of the FP-DDM algorithm rather than a validated production-scale simulator. It states that a smoke checkpoint only verifies that the execution path works, and that neural-solver accuracy must be established through a sustained training run and reported together with its exact configuration.
Scale constraints are equally explicit. Extending to very large global problems requires distributed ownership of subdomains and interfaces, whereas the current example batches local solves on a single process and retains assembled global fields for diagnostics. The synthetic global layout currently requires a square patch grid with square patches. The supplied workload uses zero heat source, matching the original FP-DDM setup. Full-domain reference solves are opt-in and limited to at most 26 subdomains; larger requests emit a warning and continue without reference metrics.
Several details directly affect reuse. Training writes a resumable state to the latest checkpoint directory and the model with the lowest isolated-patch validation loss to the best checkpoint directory, but the documentation warns that the latter is not necessarily the best checkpoint for a recursive Schwarz rollout, so candidate checkpoints must be validated on representative decompositions. Test-time adaptation is disabled by default and must be enabled explicitly through a step count parameter. SciPy and OpenSimplex provide the smooth synthetic material layouts, and Matplotlib with ImageIO are needed only for visual output.
At the physics level the assumptions are narrow. The thermal case is steady, two-dimensional, and purely conductive, with no advection, no radiative coupling, and no transient term. The elasticity comparison is a manufactured-solution problem used to validate the decomposition workflow itself. Device electrical physics, including carrier transport, interface states, and doping-dependent transport models, is entirely outside this code path.
It is therefore inaccurate to describe this chain as a full TCAD surrogate. It currently covers one segment of chip-scale thermal and mechanical stress analysis, using a neural operator in place of a local thermal solver inside a decomposition framework.
Getting Started
The official NVIDIA container is the shortest route to a working environment, since it ships with PyTorch and CUDA dependencies already installed. Clone the repository and install the example requirements from the requirements file under the fp_ddm example directory.
Training an FNO uses the train entry point directly. Configuration is managed by Hydra, and every setting can be overridden from the command line. Distributed training reuses the same entry point with torchrun, for example with a standalone launch and a node process count of four, pointing the output directory at a training output path. Resume from the latest checkpoint by passing its path as the resume directory.
The reference implementation recommends running the numerical path before the neural path, since the numerical run validates the decomposition workflow and the metric recording rather than model quality. First run with the fem solver, specify the subdomain grid rows and columns, set a modest iteration limit such as ten, disable visualisation, and give an output directory. Then switch to the fno solver, point the checkpoint directory at the best checkpoint produced by training, raise the iteration limit to around fifty, and keep visualisation off. Physics-guided test-time adaptation is enabled by adding a step count parameter; without it the checkpoint is used directly.
The elasticity comparison runs as its own script, printing relative displacement and stress errors and writing a three-panel figure with the monolithic result, the decomposition result, and their absolute difference in displacement magnitude and von Mises stress. Pass a no-plot flag when only the numbers are needed.
Each run writes the resolved Hydra configuration, per-iteration metrics, and NumPy reference fields when requested. Visualisation additionally writes a static conductivity image, temperature images, an iteration series, and an MP4 animation.
One engineering distinction deserves emphasis. Distributed training and distributed solving are separate concerns. The distributed launcher addresses parallelism on the training side only. The reference implementation states that large-domain solving needs distributed ownership of subdomains and interfaces, which it does not currently provide, so making the training script distributed does not confer distributed solving capability.
Comparison with Alternatives
Placing FP-DDM against the wider physics AI toolchain clarifies what is distinctive about it.
The first comparison is whole-domain FNO regression, the path covered by tools/neuraloperator. Whole-domain operators perform well on regular grids, but for chip-scale problems with very large domains and strong dependence on partitioning, memory and training costs become prohibitive. FP-DDM circumvents that bottleneck through decomposition, at the cost of introducing the interface-consistency problem.
The second comparison is classical Schwarz domain decomposition, which is exactly what the fem path implements. FP-DDM layers a neural local solver and a foundation-model-style physics-guided adaptation on top of that. The presence of the numerical path carries engineering value of its own, because it serves as the reference baseline against which the neural solution is judged.
The third comparison is the X-MeshGraphNet reservoir example covered in cases/physicsnemo-reservoir-simulation. That route follows graph neural networks over finite-volume reservoir data and handles irregular connectivity such as faults and pinch-outs. FP-DDM follows the operator route with explicit interface optimisation, which suits problems with a well-defined equation structure whose residual can be evaluated. The distinction can be summarised as follows: the reservoir surrogate learns how to propagate state across a cell graph, whereas FP-DDM learns how to align solutions at subdomain boundaries.
The fourth comparison is the SK hynix etch surrogate. It is based on a graph network simulator that combines numerical time stepping with a graph neural network to model geometry evolving over time, which suits a process where the geometry changes continuously. The rationale given for this architecture is that neural operators typically need large datasets and struggle under data scarcity, whereas the graph simulator captures local interactions and represents critical physical properties with minimal training data.
The fifth comparison is a purely data-driven graph surrogate, which is where the SK hynix effort started with MeshGraphNet. FP-DDM instead trains with physics information, an FNO fitted without solution labels using residuals as the loss, placing it closer to the physical-constrained learning discussed in collections/pinn-getting-started.
The engineering trade-off is now visible. The operator route suits problems with known equations and computable residuals. The graph route suits problems where state propagates over a discrete graph and geometry evolves dynamically.
Benchmarks and Evidence
The FP-DDM paper appears in the DAC 2026 conference proceedings under the title Neural Domain Decomposition for Scalable Multi-Physics: Chip-Scale Thermal-Stress Analysis. The author list spans Min-Chul Park, Ji-Hye Lee, Sungyeop Lee, Giyong Hong, Seokki Lee, In Huh, Hong-hyun Park, Hyunjae Jang, Changwook Jeong, Young-Gu Kim and Dae Sin Kim, with affiliations across Samsung Electronics, Samsung Semiconductor, UNIST and Ulsan National Institute of Science and Technology, and a DOI of 10.1145/3770743.3803918. The reported magnitudes are solver-level accuracy on thermal and stress problems, extension to domains with billions of cells, and support for layout design optimisation.
Two technical points in the conference abstract are worth recording. The first concerns how interface consistency is achieved, namely by partitioning into overlapping subdomains and explicitly optimising boundary values over the overlap through automatic differentiation for fast alignment. The second concerns generalisation, namely a foundation architecture that learns multi-physics behaviour and transfers to new domains, combined with foundation-model adaptation and physics-guided test-time adaptation to reach unseen physics and unseen domains without labels.
The reference implementation ships no public benchmark dataset. It generates local thermal problems on the fly and constructs smooth synthetic material layouts with SciPy and OpenSimplex, so what can be reproduced is the workflow, the metrics and the interface, not a table of numbers matching the paper. That distinction matters when citing results, since the accuracy claims belong to the paper while the runnable artefacts belong to the repository.
The SK hynix case is documented in terms of methodology rather than error budgets. The NVIDIA technical blog titled Using AI Physics for Technology Computer-Aided Design Simulations was published on 17 December 2025 by Ram Cherukuri, Kihang Youn, Gyuseung Han, Min Kang, Hwiwon Seo and Junghan Kim, with contributions from Yiyi Wang and Alexey Kamenev. Its table lists six successive improvements with their effect: adopting MeshGraphNet decreased memory requirements; using Chamfer loss for velocity calculation reduced training loss; re-meshing at each iteration improved inference accuracy; feature selection improved inference accuracy; multi-scale message passing reduced training loss; and updating material features at each iteration improved inference accuracy.
At the level of industrial effect the statement is qualitative. The team argues that AI-augmented TCAD will become a key enabler of research productivity in the semiconductor industry, allowing engineers to evaluate tens of thousands of process cases generated from dozens of recipe combinations and turning TCAD from qualitative guidance into a quantitative optimisation framework.
The two bodies of evidence differ in kind. The DAC paper supplies scale and accuracy magnitudes, while the SK hynix blog supplies a reusable sequence of methodological refinements. Neither provides a general error budget that can be transferred directly to another flow.
Role in Physics-AI Workflows
Within a semiconductor process, surrogate models land in several places, and the density of value differs considerably between them.
The foremost is rapid scanning of layout and structural design. The paper explicitly mentions support for layout design optimisation, which is what distinguishes FP-DDM from an ordinary PDE surrogate. The target is not single-point solving but placing design variables such as layout, conductivity distribution, and heat source position inside the decomposition framework for iterative exploration. Running a full-chip solve for every candidate layout is close to unaffordable.
The second is morphology surrogates in the process direction. For processes such as etching where geometry evolves over time, parameter sweeps are reorganised in the deep learning era into replacing process simulation with a surrogate. The SK hynix path starts from a graph network simulator that captures local interactions, then improves morphology prediction accuracy through loss design and iteration strategy, with Chamfer loss for the velocity field, re-meshing for inference stability, and material feature updates each iteration to capture material state evolution.
The third is joint evaluation of thermal and mechanical reliability. Heat and stress are not independent problems in chip design, since thermal gradients produce differential thermal expansion that is a primary source of stress. The decomposition framework is naturally suited to such multi-physics coupling, because each subdomain solves the coupled equations internally while interface consistency is guaranteed by a single explicit optimisation mechanism.
The general physics AI recipe is fully visible in this chain. Physics-informed training addresses data scarcity. Decomposition addresses scale. Test-time adaptation addresses transfer across domains. A numerical reference baseline addresses credibility. Each element answers a distinct failure mode, and the combination is more complete than any single technique.
Its relation to other entries on this site is complementary. cases/openfoam-multiphysics covers the boundary conditions of a general multi-physics CFD pipeline, and cases/battery-thermal-piml covers surrogate modelling for coupled electrochemical and thermal problems. Together with this entry they form three segments of one thread around multi-physics plus surrogates.
Limitations and Failure Modes
Some exclusions are absolute. Device electrical problems are out of scope. Drift-diffusion transport, carrier statistics, interface traps and Fermi-level pinning, and the influence of doping profiles on junction characteristics are core device TCAD content, whereas the FP-DDM reference implementation handles heat conduction plus an elasticity baseline. Applying this chain to device-level electrical surrogacy uses the wrong instrument.
Three-dimensional problems are also outside the current implementation. The example is a two-dimensional thermal equation with a two-dimensional plane-stress comparison. Although the paper describes a billion-cell scale, that reflects the implementation described in the paper rather than a capability offered by the repository reference implementation. The synthetic global layout additionally requires a square patch grid with square blocks, which restricts direct application to irregular chip outlines.
Several concrete failure modes follow from that. The most common is mis-selected checkpoints. The best checkpoint is chosen by isolated-patch validation loss, and the documentation explicitly warns this is not necessarily the best checkpoint for a recursive Schwarz rollout. Ignoring this leads the neural path to diverge or stall after several iterations, which is easily misdiagnosed as the method being unsuitable.
A second is treating an exhausted iteration budget as convergence. The stopping metric is overlap RMSE, and the documentation is explicit that reaching the maximum iteration count is only a budget limit. Ignoring residual or interface-consistency trends and merely checking whether the iteration count completed yields a falsely stable conclusion.
A third is overestimating the benefit of test-time adaptation. The adaptation step count is not set by default and must be supplied explicitly. It continues to optimise the FNO on the current subdomain batch, and its value depends on the physical diversity of that batch. Where the data distribution is narrow, adaptation may merely pull the model toward the current batch without improving generalisation to unseen domains.
There is also a gap between the physics residual and the true solution. The FNO is trained on a combination of thermal and boundary residuals, and a small residual does not necessarily imply good interface consistency or engineering accuracy. Where conductivity differs sharply between domains, the linear interpolation assumed across the overlap introduces systematic bias that may remain invisible in residual metrics.
Finally, confusing scale with parallelism. Treating distributed training as conferring distributed solving capability is the easiest engineering error on this chain, and the reference implementation states plainly that large-domain solving requires distributed ownership of subdomains and interfaces, which the current version does not provide.
Version and Source Notes
This section records each correction made to the earlier wording of this entry together with its basis, so that citations can be checked against the original claim.
First, the actual content of the example directory. The earlier entry described the PhysicsNeMo TCAD examples as replacing repeated device-level numerical solves with data-driven models, taking geometry and doping profiles plus bias conditions as input and returning field or current distributions as output. The tcad directory currently holds a single sub-example, fp_ddm, which handles the steady two-dimensional heat equation and a two-dimensional plane-stress elasticity comparison, with synthetic conductivity distributions, a volumetric heat source and temperature boundaries as input and a temperature field plus displacement and stress errors as output. Doping profiles, bias conditions and current distributions are not part of it.
Second, the scenario framing. The earlier entry wrote that the work targets thermal-electric co-design for 3D-IC and advanced packaging. The reference implementation is a steady two-dimensional thermal problem plus a two-dimensional plane-stress elasticity comparison, which is thermal-mechanical coupling in the thermal-stress direction, with no electrical coupling included. Three-dimensional thermal-electric analysis at 3D-IC and advanced packaging scale is not on this code path. The claims about three dimensions and chip-scale layout optimisation come from the DAC 2026 method description and are capability statements at the paper level.
Third, the distinction between etching and device electronics. The earlier entry mentioned no industrial case and did not separate process TCAD from device TCAD. The industrial case that is on the record is the etch process surrogate from the TCAD intelligence team at SK hynix, which belongs to process TCAD. It is a separate technical path from device-level electrical surrogacy, and the two should not be merged in citations.
Fourth, the composition of TCAD itself. The NVIDIA blog splits TCAD into process simulation of physical and chemical steps such as deposition, lithography, etching and ion implantation, and device simulation that takes the three-dimensional structure predicted by the process and models its electrical behaviour. The earlier wording referred only to device-level drift-diffusion or Monte-Carlo transport and omitted the process side, which is where the actual industrial case sits.
Fifth, the positioning of other architectures. The NVIDIA blog describes a generic onboarding path: reproduce the MeshGraphNet training pipeline from the vortex shedding example, customise it on your own data, optionally evaluate architectures such as DoMINO or Transolver, and use the built-in distributed capability to scale to three-dimensional chip-scale simulations. That is framework guidance rather than the same workflow as the fp_ddm reference implementation, and conflating the two leads to misjudging the example's capabilities.
Sixth, the citation boundary of accuracy claims. The statements that FP-DDM reaches solver-level accuracy and scales to billions of cells are paper conclusions and should be attributed to the DAC 2026 paper. The repository describes itself as a research example rather than a production-grade simulator, with accuracy to be established by sustained training and reported alongside its configuration. The two cannot be cited interchangeably.
Seventh, data availability. The reference implementation generates synthetic local thermal problems on the fly and ships no public benchmark dataset. What can be reproduced is the workflow, the metrics and the interface rather than the paper's numerical results. Accuracy figures should be attributed to the paper rather than to the repository example.
Eighth, checkpoint semantics. The best checkpoint is selected by isolated-patch validation loss, and the documentation states that it is not necessarily the best checkpoint for a recursive Schwarz rollout. This is a necessary evaluation caveat that the earlier entry did not address at all.
Ninth, the stopping criterion. Overlap RMSE is the stopping metric, and reaching the iteration ceiling is a budget limit rather than convergence. The earlier entry contained no iteration-level criterion, so the expanded version must add one to avoid presenting a non-converged rollout as a result.
Tenth, the wording used for engineering maturity. The reference implementation requires distributed ownership of subdomains and interfaces in order to extend to large domains, and currently batches local solves on a single process. The expanded text consistently uses the description of a research example rather than claiming production readiness.
Eleventh, year and licence. The year field moves from 2023 to 2026, since the FP-DDM path and the DAC 2026 paper landed together and the earlier value matches none of the empirical sources cited here. The PhysicsNeMo repository is released under Apache-2.0 and the DAC 2026 paper copyright is held by its authors, so the licence field separates the two.
References
- NVIDIA PhysicsNeMo reference implementation and README for
examples/tcad/fp_ddm, https://github.com/NVIDIA/physicsnemo/tree/main/examples/tcad/fp_ddm - Park M.-C., Lee J.-H., Lee S., Hong G., Lee S., Huh I., Park H.-H., Jang H., Jeong C., Kim Y.-G., Kim D. S., "Neural Domain Decomposition for Scalable Multi-Physics: Chip-Scale Thermal-Stress Analysis", Proceedings of the 63rd ACM/IEEE Design Automation Conference (DAC '26), 2026, DOI 10.1145/3770743.3803918
- DAC 2026 conference program abstract page, session "Hot Chips & Cool Models: AI Turns Up the Heat on Silicon Design", https://63dac.conference-program.com/presentation/?id=RESEARCH158
- Cherukuri R., Youn K., Han G., Kang M., Seo H., Kim J., "Using AI Physics for Technology Computer-Aided Design Simulations", NVIDIA Technical Blog, 2025-12-17, https://developer.nvidia.com/blog/using-ai-physics-for-technology-computer-aided-design-simulations/
- NVIDIA PhysicsNeMo examples overview
examples/README.md, including the architecture comparison table for TCAD and reservoir simulation, https://github.com/NVIDIA/physicsnemo/blob/main/examples/README.md - Li Z., Kovachki J., Azizzadenesheli K., Liu J., Bhattacharya K., Stuart A., Anandkumar A., "Fourier Neural Operator for Parametric Partial Differential Equations", arXiv:2010.08895, 2020
- Li Z., Kovachki J., Azizzadenesheli K., Liu J., Bhattacharya K., Stuart A., Anandkumar A., "Physics-Informed Neural Operator for Learning Partial Differential Equations", arXiv:2111.03794, 2021
- NVIDIA PhysicsNeMo documentation site covering installation, user guide and API reference, https://docs.nvidia.com/physicsnemo/latest/index.html
- Official NVIDIA PhysicsNeMo container catalogue on NGC, https://catalog.ngc.nvidia.com/orgs/nvidia/teams/physicsnemo/containers/physicsnemo
- NVIDIA announcement of open-source models and the Apollo pre-trained model set presented at SC25, https://blogs.nvidia.com/blog/apollo-open-models/
- Nabian M. A., Liu C., Ranade R., Choudhry S., "X-MeshGraphNet: Scalable Multi-Scale Graph Neural Networks for Physics Simulation", arXiv:2411.17164, 2024
- NVIDIA PhysicsNeMo vortex shedding example, referenced as the onboarding path for TCAD surrogate development, https://github.com/NVIDIA/physicsnemo/tree/main/examples/cfd/vortex_shedding_mgn