WorksComputer Science
IC-ThermBench: An Open, Progressive Benchmark for Generalizable 2.5D/3D-IC Thermal Learning
Abstract
Standardized benchmarks are fundamental to reliable progress in AI for EDA, including learning-based thermal modeling. However, existing thermal prediction studies often rely on different datasets, simulators, data splits, preprocessing pipelines, and metrics, while most datasets and implementations remain unavailable, making fair and reproducible comparison difficult. We introduce IC-ThermBench, an open and progressive benchmark that combines established 3D-IC steady-state, transient, and industrial package tasks with a new 50,000-sample 2.5D chiplet extension designed to evaluate progressively broader represented physical variation and cross-package OOD transfer. Five Generalization Scopes cover 3D-IC fixed-design prediction, Within-Family Generalization under layout, material, and boundary-condition variation, and Cross-Package OOD transfer to unseen package systems. We evaluate eight representative baselines under common data, splits, labels, and metrics. Performance degrades gradually from S2 to S4 as represented physical support broadens, but Cross-Package OOD produces a much sharper degradation: the best RMSE and MAE increase from 1.216 and 0.938~K at S4 to 15.99 and 15.00~K at S5, respectively. With only 10 labeled samples per OOD case, target-domain adaptation reduces the best MAE to 2.60~K. IC-ThermBench further provides a unified generation, training, inference, and evaluation pipeline, enabling reproducible and fair comparison of existing and new thermal predictor.
Keywords thermal benchmark, public dataset, chiplets, 2.5D/3D integration, neural operators, out-of-distribution generalization
Introduction
Thermal analysis is critical to 2.5D/3D-IC design, where chiplet placement, power density, material properties, and cooling conditions jointly determine temperature distributions and hotspots. Repeated high-fidelity simulation during placement, routing, and design-space exploration can be expensive, motivating learning-based thermal surrogates. Learning-based thermal predictors have recently attracted increasing attention, achieving high predictive accuracy while enabling fast inference for repeated thermal analysis [9, 16, 32, 33].
Prediction accuracy alone, however, does not make models directly comparable. Benchmarks such as ImageNet, SWE-bench, and PDEBench, together with CircuitNet in EDA, demonstrate the importance of shared data, task definitions, splits, metrics, and evaluation procedures [1, 5, 10, 25]. This is particularly important for thermal learning, where differences in datasets, simulators, preprocessing, and test distributions can substantially affect reported errors. Results obtained under different evaluation settings therefore cannot be interpreted as model-to-model improvements by default.
Existing thermal-learning studies have expanded beyond power-only variation to more challenging geometry, material, and cooling conditions [18, 34, 37, 38]. However, their datasets, representations, splits, and evaluation protocols remain highly heterogeneous, and complete data and code are often unavailable. Table 1 summarizes these differences; the Closest-Scope labels indicate conceptual similarity rather than protocol equivalence. The field therefore still lacks an open and unified framework for fair and reproducible comparison across thermal predictors.
| Work | Publication | Model family | Labels / scale | Factors | Generalization / evaluation setting | Closest Scope | Data public | Code public |
| ThermGAN [11] | ICCAD’20 | conditional GAN | commercial CPU traces; NR | workload-conditioned maps on fixed CPU designs | S1 | ✗ | ✗ | |
| ThermEDGe [3] | ASP-DAC’21 | encoder–decoder / ConvLSTM | private numerical; 5k | within-design steady/transient thermal and IR tasks | S1 | ✗ | ✗ | |
| GCN+PNA [2] | ASP-DAC’22 | graph convolution | HotSpot; six unseen sets | inductive evaluation across unseen chiplet layouts/sizes | S2/S5 | ✗ | ✗ | |
| DeepOHeat [16] | DAC’23 | DeepONet operator | PDE residuals; Celsius check | parametric/non-parametric configurations; no supervised labels in original method | S4 | – | ✓ | |
| MCM-DCNN [6] | TSEP’23 | SDF-encoded DCNN | numerical; NR | trains on 2–4 chips and tests a 5-chip configuration | S2/S5 | ✗ | ✗ | |
| ARO [28] | ICCAD’24 | autoregressive operator | MTA [12]; 3×5k | within-design power; cross-case transfer/multi-fidelity | S1 | ✓ | ✗ | |
| FaSTherm [39] | ICCAD’24 | residual CNN | full-chip transient; NR | nonlinear transient prediction on task-specific designs | S1 | ✗ | ✗ | |
| T-Fusion [36] | ASP-DAC’25 | Bayesian tensor fusion | EV6 MTA; HF/LF subsets | fixed-design transient multi-fidelity fusion | S1 | ✗ | ✗ | |
| SAU-FNO [7] | DAC’25 | attention U-FNO | MTA; 3×5k | fixed-design power; transfer between source cases | S1 | ✗ | ✗ | |
| FSA-Heat [38] | DATE’25 | frequency–spatial net | HotSpot; 6k | represented mixtures plus unseen conductivity/source count | S3/S5 | ✗ | ✗ | |
| Therunet-SP [13] | ISQED’25 | scaled U-Net | 2.5D numerical; NR | geometry scaling across 2.5D configurations | S2 | ✗ | ✗ | |
| Therm-PCT [37] | ICCAD’25 | point-cloud Transformer | COMSOL point clouds; 8.5k | unseen shapes and point-cloud sizes, including larger unstructured domains | S3/S5 | ✗ | ✗ | |
| Adaptive Graph [34] | ICCAD’25 | GNN–FEM hybrid | fine-grid numerical + silicon; NR | variable structure/interface; unseen process/material without retraining | S3/S5 | ✗ | ✗ | |
| PNO-Therm [8] | DAC’26 | PDE foundation model | HS source suite; 3×5k | fine-tuning on fixed-design steady tasks | S1 | ✗ | ✗ | |
| COOL [18] | DAC’26 | 3D point Transformer | commercial FEM; 74×10 | design-disjoint 58/16 split over heterogeneous 3D/3.5D systems | S4/S5 | ✗ | ✗ |
Table 1. Expanded prior-work survey of learned thermal predictors. These entries are not the roster of evaluated IC-ThermBench baselines. Labels/scale identifies the reference data source, not the predictor. ✓/✗ mark whether data and code are public; “–” marks DeepOHeat data as not applicable because its original training is label-free.
To fill this gap, we introduce IC-ThermBench, an open and progressive benchmark for 2.5D/3D-IC thermal learning. It combines established 3D-IC steady-state, transient, and industrial package tasks with a new 50,000-sample 2.5D chiplet extension that systematically broadens the represented physical variation. Five Generalization Scopes span increasingly challenging deployment settings: S1 preserves established 3D-IC fixed-design tasks; S2–S4 progressively introduce layout, material, and boundary-condition variation within represented system families; and S5 evaluates case-disjoint transfer to previously unseen package systems. This organization explicitly separates Within-Family Generalization from Cross-Package OOD transfer.
The main contributions are threefold:
We establish an open thermal-learning benchmark that unifies established public tasks with a new 50,000-sample progressive extension under a consistent data and evaluation framework.
We benchmark eight classic and state-of-the-art thermal predictors under a unified protocol. Results show moderate degradation with broader represented physical variation, a sharp drop under Cross-Package OOD, and substantial recovery with limited target-domain adaptation.
We provide an end-to-end benchmark toolkit covering data generation, format conversion, immutable splits, training,
inference, and evaluation, allowing existing and new models to be integrated, reproduced, and fairly compared through a common interface. Models [7, 9, 14, 16, 17, 20, 23, 28, 31, 35]. These approaches have achieved high prediction accuracy across a variety of fixed-design or represented-distribution settings.
Related Work
Benchmarks in ML, Scientific ML, and EDA
Modern benchmarks provide not only datasets but also shared task definitions, data splits, metrics, and evaluation protocols that enable reproducible and fair model comparison. ImageNet, GLUE, SWE-bench, and PDEBench established this paradigm across vision, language, software agents, and scientific ML [5, 10, 25, 27]. Learning-based EDA has likewise benefited from public benchmarks such as CircuitNet, AMSNet, OpenABC-D, EDALearn, and ChiPBench, which provide reusable foundations for model development and evaluation [1, 4, 19, 26, 30].
Such standardization is particularly important for thermal learning, where differences in simulation settings, physical conditions, data splits, input representations, and evaluation metrics can substantially affect reported errors and make independently reported results difficult to compare directly.
Learning-Based Thermal Simulation
Learning-based thermal simulation has explored convolutional and image-to-image models [3, 6, 11, 15, 21, 24, 39], as well as neural operators, physics-informed learning, and pretrained PDE models [7, 9, 14, 16, 17, 20, 23, 28, 31, 35]. These approaches have achieved high prediction accuracy across a variety of fixed-design or represented-distribution settings.
Recent studies further extend thermal prediction toward cross-configuration and cross-system generalization. Therunet-SP studies geometry scaling, ThermGCN evaluates unseen chiplet layouts, FSA-Heat jointly varies geometry, power, material, and heat-exchange conditions, Therm-PCT targets unseen unstructured geometries, and Adaptive Graph Learning considers process, material, and interface variation [2, 13, 34, 37, 38]. COOL further represents geometry, power, material, and cooling conditions in a unified formulation and evaluates generalization with a design-disjoint split [18]. Alpha EV6-based tasks have also been adopted by multiple studies, including ARO, T-Fusion, SAU-FNO, and Therm-FM [7, 9, 28, 36].
Despite this progress, evaluation settings in thermal learning remain highly fragmented. Different studies often use different datasets, simulators, input representations, data splits, and metrics, while complete data and implementations are frequently unavailable. As a result, existing results are difficult to reproduce and compare fairly, and evaluating new methods often requires rebuilding study-specific data and evaluation pipelines.
Problem Formulation
For a packaged chiplet system, the temperature field is jointly determined by the system structure and its physical operating conditions. Let denote the structural configuration, including the package outline, layer stack, and chiplet organization, and let collect the physical variables exposed by a task, such as power, material properties, boundary conditions, and, when applicable, time. The underlying thermal process follows the heat equation,
and can be abstracted as a thermal solution operator
For steady-state analysis, the temporal term vanishes; transient tasks instead predict the temperature evolution over a specified time horizon. S1 retains the source definitions of its steady-state and transient labels, whereas S2–S5 use HotSpot-generated steady-state labels.
Each benchmark sample is represented as , where contains the physical variables exposed by the corresponding task and is the reference temperature field. Models may differ in architecture and optimization, while the benchmark fixes the data, splits, prediction targets, and evaluation metrics within each track.
Benchmark Design and Protocol
Benchmark Scope and Progressive Design
IC-ThermBench organizes thermal prediction into five deployment-oriented Generalization Scopes, rather than a universal difficulty ranking. S1 evaluates within-design prediction; S2–S4 progressively broaden the represented physical support over the same observed system families through layout, material, and boundary-condition variation; and S5 evaluates case-disjoint transfer to previously unseen package systems.
The distinction between S2–S4 and S5 is central to the benchmark. S2–S4 are independently sampled and therefore represent progressively broader training distributions rather than paired one-factor interventions. S5 instead changes the underlying package families, which may jointly shift structural properties and the empirical distribution of model inputs. We therefore call S2–S4 Within-Family Generalization: test samples come from package/system families represented during training, while their physical conditions broaden. S5 is Cross-Package OOD: all test package systems are unseen during training. Figure 1 summarizes this organization.

Figure 1. Overview of IC-ThermBench. S1 preserves established within-design source tasks. S2–S4 evaluate Within-Family Generalization under layout, material, and boundary-condition variation over package/system families represented during training. S5 evaluates Cross-Package OOD on package systems unseen during training.
Five Generalization Scopes
Figure 1 summarizes the systems, inputs, sample counts, and primary variation of the five Scopes.
S1: established source suite. S1 consolidates the established Therm-FM source suite [9], including Alpha EV6 steady-state/transient tasks and industrial package tasks. It preserves the source task definitions and fidelity settings while using the split fixed by the current benchmark release. It serves as the benchmark’s within-design reference.
S2–S4: Within-Family Generalization. S2–S4 share the same ten ATPplace2.5D system families [29] while progressively broadening the represented physical variation. S2 varies chiplet placement and orientation; S3 additionally varies filler thermal conductivity; and S4 further varies ambient temperature and convective cooling. The three datasets are independently sampled and therefore characterize progressively broader represented distributions rather than sample-wise paired ablations.
S5: Cross-Package OOD. S5 contains five held-out package systems, Cases 11–15, that do not participate in S2–S4 training or validation. They primarily stress chiplet count, power density, size heterogeneity, power concentration, and package utilization. For example, Case 11 increases the chiplet count from a training maximum of 61 to 80, while Case 14 concentrates 90.9% of total power in one chiplet compared with a training maximum of 69.2%. The remaining held-out cases primarily stress power density, size heterogeneity, and utilization. The five held-out systems and their stress characteristics were fixed before baseline evaluation and are shared by every model and metric.
Data Generation and Input Representation
For S2–S5, feasible chiplet layouts are rasterized onto a grid and simulated with the open-source HotSpot thermal simulator using its grid-based detailed-3D flow. All generated tasks are steady-state, with temperature fields stored in kelvin as prediction targets. S1 instead preserves the source task definitions and fidelity settings while using the split fixed by the benchmark release.
The model input expands cumulatively with the Generalization Scope. Let denote the rasterized nominal chiplet-power field, and the physical coordinates, the local thermal-conductivity field, and , , and the package-level boundary quantities. The resulting inputs are
The power channel broadcasts each chiplet’s nominal power over the grid cells occupied by that chiplet and is therefore not a pixel-wise power-density field. The and channels encode physical coordinates in millimeters. Layouts use random, dispersed, cluster, and boundary modes in a 6:2:1:1 ratio. For S3 and above, filler thermal conductivity is sampled log-uniformly from W/(m·K), while chiplet silicon remains fixed at 100 W/(m·K). For S4 and S5, is sampled uniformly from °C and log-uniformly from W/(mK); is derived from and heatsink geometry rather than independently sampled. Complete tensor semantics and HotSpot configurations are provided with the benchmark release.
Evaluation Protocol and Metrics
For S2–S4, each case contributes 1,080/120/300 training/validation/test samples, yielding 10,800/1,200/3,000 samples per Scope. The split is case-balanced, and all normalization statistics are computed from the training subset only. Because Cases 1–10 remain represented across the split, these Scopes evaluate Within-Family Generalization under progressively broader represented physical variation.
The primary S5 protocol evaluates frozen S4 checkpoints directly on held-out Cases 11–15. These systems are excluded from training, validation, normalization fitting, and checkpoint selection, and the preprocessing pipeline remains unchanged at test time. We additionally report target-domain adaptation as an auxiliary sample-efficiency track. For each S5 case, the first 500 released samples form a fixed adaptation pool and the remaining 500 form a fixed holdout. The benchmark supplies fixed nested index sets for , identical for every model, amounting to target labels in total. The adaptation baseline evaluates the frozen S4 checkpoint on this 2,500-sample holdout and is therefore distinct from the official S5 zero-shot result computed over all 5,000 samples. Normalization statistics remain frozen throughout adaptation.
We report RMSE and MAE for global field accuracy, for explained variation, MaxAE for the largest local error, and and Top-50 MAE for hotspot prediction. Except for , which is pooled over all test grid points, every metric is computed per sample and then averaged over test samples (Appendix C.4); MaxAE is thus the mean per-sample worst-pixel error. All models are compared using the same data splits, prediction targets, and evaluation metrics.
Experimental Evaluation
Experimental Setup and Baselines
We evaluate eight configurations spanning three model families: U-Net as a convolutional predictor; FNO, U-FNO, SAU-FNO, and DeepOHeat as neural-operator or operator-style models; and Therm-FM T/B/L as pretrained PDE foundation models. S1 follows the source task definitions and fidelity settings with the released benchmark split, whereas S2–S5 use the same IC-ThermBench datasets, splits, labels, input semantics, normalization, and metrics.
Fairness is enforced at the data and evaluation interfaces rather than by forcing identical optimization across heterogeneous architectures. FNO, U-FNO, SAU-FNO, U-Net, and DeepOHeat use Adam with learning rate , batch size 20, and 100 training epochs. Therm-FM uses learning rate , batch size 40, and 100 epochs. All normalization statistics are fitted using training data only and remain fixed during validation, testing, and S5 zero-shot evaluation.
All S2–S5 experiments use NVIDIA RTX A6000 48GB GPUs. When complete public training code is unavailable, we construct a paper-based reference implementation from the published architecture and available configurations and report the resulting values as reproductions rather than author-reported results. Full optimizer schedules, software versions, and implementation details are provided with the benchmark release.
| Model | Key configuration | Params |
| U-Net [22] | channels 58/116/232/348; three downsampling stages | 5.06M |
| FNO [14] | FNO3d; modes (12, 12, 1); width 72 | 11.98M |
| U-FNO [31] | Net3d; modes (10, 10, 1); width 36 | 5.10M |
| SAU-FNO [7] | attention-augmented U-FNO; modes (10, 10, 1); width 36 | 5.11M |
| DeepOHeat [16] | 128-D coordinate trunk; 256-wide 7-layer branch | 7.92M |
| Therm-FM T/B/L [9] | pretrained scOT/Poseidon checkpoints | 21/158/629M |
Table 2. Architecture-defining configurations used in the controlled S2–S5 experiments. Parameter counts correspond to the S4/S5 seven-channel configuration where applicable.
Fixed-Design Reference and Within-Family Generalization (S1–S4)
S1 serves as a within-design performance anchor under the source Alpha EV6 and industrial-package task definitions and fidelity settings. Figure 2 presents this fixed-design anchor alongside the S2–S4 Within-Family Generalization progression. Across S2–S4, MAE and RMSE show the same gradual degradation as represented physical variation broadens. Table 3 summarizes five representative baselines at the finest reported resolution. Across convolutional/operator-style predictors, prediction errors are generally low when geometry, materials, and cooling conditions remain fixed, while the pretrained Therm-FM-L further reduces MAE across all three task groups. This establishes a within-design reference before progressively broadening the physical support in S2–S4.
S2–S4 evaluate increasingly broad physical variability while retaining the represented Cases 1–10 system families. Because the three datasets are independently generated rather than sample-wise paired, their comparison reflects increasing distributional difficulty rather than a causal ablation of individual variables. Table 4 reports global-field RMSE and MAE together with worst-pixel MaxAE and hotspot-region errors. For context, the same metrics are also shown for the S5 Cross-Package OOD evaluation.

Figure 2. MAE and RMSE from the S1 fixed-design anchor through S2–S4 Within-Family Generalization. All eight baselines, including Therm-FM T/B/L, are retained across the four settings. S2–S4 progressively broaden represented layout, material, and boundary-condition variation. Both panels use the same logarithmic error axis.
| Method | HS steady | HS transient | Industrial |
| FNO | 0.063–0.164 | 0.034–0.163 | 0.017–0.018 |
| U-FNO | 0.046–0.126 | 0.011–0.144 | 0.013 |
| SAU-FNO | 0.041–0.162 | 0.013–0.149 | 0.020–0.026 |
| DeepOHeat | 0.745–2.585 | 0.315–1.820 | 0.039–0.044 |
| Therm-FM-L (629M) | 0.012–0.049 | 0.004–0.030 | 0.008 |
Table 3. Compact S1 source-suite summary at the finest reported resolution. Values are MAE ranges in kelvin across each task group; results are not pooled across systems or resolutions. Therm-FM values coincide with those reported in the Therm-FM study [9]; the other baselines are trained and evaluated in the current benchmark.
Across the represented system families, the best RMSE increases from 0.354 K at S2 to 0.670 K at S3 and 0.811 K at S4. The best MAE follows the same gradual trend, increasing from 0.248 to 0.493 and 0.585 K, respectively. Thus, the broader S3 and S4 distributions degrade global-field accuracy progressively rather than producing an abrupt failure.
Model ordering also changes as the represented physical support expands. Therm-FM-L is numerically best on all four reported metrics at S2 and S4, whereas U-FNO leads all four at S3, ahead of every Therm-FM variant; U-FNO also stays within roughly 1.26× of the best RMSE on every within-family Scope. Scaling Therm-FM improves accuracy monotonically at S2 and S4 but not at S3, where Therm-FM-B slightly outperforms Therm-FM-L. S4 therefore provides the within-family checkpoint from which we next evaluate transfer to unseen package families.
Cross-Package OOD and Adaptation (S5)
S5 evaluates whether the within-family performance observed in S2–S4 transfers to previously unseen package systems. We directly evaluate frozen S4 checkpoints on Cases 11–15 without target-domain labels or updated normalization statistics. Unlike the gradual degradation from S2 to S4, cross-package transfer produces a sharp performance break. Figure 3(a) shows the zero-shot RMSE, MAE, and MaxAE, while Table 4 reports the complete condensed comparison: the best RMSE increases from 0.811 K at S4 to 8.67 K at S5, approximately a 10.7× increase, while the best MAE rises from 0.585 to 8.00 K, approximately a 13.7× increase.
| Method | S2: Layout RMSE | S2: Layout MAE | S2: Layout MaxAE | S2: Layout Top50 | S3: Material RMSE | S3: Material MAE | S3: Material MaxAE | S3: Material Top50 | S4: Boundary RMSE | S4: Boundary MAE | S4: Boundary MaxAE | S4: Boundary Top50 | S5: Unseen Packages RMSE | S5: Unseen Packages MAE | S5: Unseen Packages MaxAE | S5: Unseen Packages Top50 |
| FNO | 0.828 | 0.621 | 3.965 | 0.494 | 0.962 | 0.720 | 4.472 | 0.506 | 1.291 | 0.994 | 5.436 | 0.693 | 42.175 | 39.358 | 65.918 | 39.713 |
| U-Net | 0.738 | 0.541 | 5.273 | 0.500 | 0.843 | 0.630 | 5.020 | 0.502 | 1.271 | 0.976 | 7.017 | 0.838 | 18.854 | 17.374 | 32.432 | 18.064 |
| U-FNO | 0.444 | 0.326 | 2.757 | 0.298 | 0.670 | 0.493 | 3.367 | 0.373 | 0.939 | 0.724 | 5.011 | 0.573 | 28.154 | 27.060 | 39.489 | 26.647 |
| SAU-FNO | 0.534 | 0.409 | 2.632 | 0.428 | 0.708 | 0.525 | 3.700 | 0.447 | 0.953 | 0.739 | 4.137 | 0.683 | 26.121 | 25.056 | 36.358 | 26.902 |
| DeepOHeat | 3.706 | 2.938 | 10.261 | 3.645 | 3.970 | 3.226 | 10.553 | 4.826 | 4.739 | 3.903 | 11.755 | 5.924 | 23.973 | 23.141 | 33.433 | 24.078 |
| Therm-FM-T | 0.575 | 0.410 | 3.121 | 0.399 | 0.911 | 0.679 | 4.268 | 0.551 | 1.050 | 0.794 | 4.843 | 0.643 | 8.673 | 8.003 | 15.535 | 9.045 |
| Therm-FM-B | 0.461 | 0.320 | 2.709 | 0.311 | 0.722 | 0.517 | 3.717 | 0.378 | 0.846 | 0.618 | 4.260 | 0.461 | 13.055 | 12.339 | 21.286 | 15.746 |
| Therm-FM-L | 0.354 | 0.248 | 2.145 | 0.238 | 0.737 | 0.527 | 3.785 | 0.390 | 0.811 | 0.585 | 4.034 | 0.431 | 15.336 | 14.633 | 24.088 | 13.768 |
Table 4. Main IC-ThermBench results under Within-Family Generalization and Cross-Package OOD. S2–S4 progressively broaden represented physical variation over the same Cases 1–10 package families, whereas S5 evaluates held-out Cases 11–15 and is therefore reported as a separate Cross-Package OOD track. MaxAE denotes the mean per-sample worst-pixel error and Top50 denotes Top-50 MAE. All errors are in kelvin and lower is better; best results within each Scope and metric are bolded.
Figure 3. Cross-Package OOD field errors before and after target-case adaptation. (a) Official S5 zero-shot results on all 5,000 held-out samples. (b) S5 10-shot adaptation, with 10 labeled samples per held-out case, evaluated on the fixed 2,500-sample adaptation holdout. Both panels show the same five conventional/operator baselines and Therm-FM-T. RMSE and MAE measure global-field accuracy, while MaxAE is the mean per-sample worst-pixel error. Both axes are zero-based, although their ranges differ; lower is better. Complete eight-model zero-shot results remain in Table 4.
The model ranking also changes substantially. U-FNO and SAU-FNO, the strongest operator baselines under S2–S4, degrade severely on S5, and Therm-FM-L—the best S2 model—becomes the weakest Therm-FM variant here in RMSE, MAE, and MaxAE, whereas Therm-FM-T is the strongest zero-shot model despite trailing B and L within family; FNO degrades most, reaching 42.18 K RMSE. Even the best S5 result remains far above the S4 within-family accuracy. The per-case RMSE profiles are also strongly case dependent: DeepOHeat ranges from 4.79 K on Case 11 to 44.63 K on Case 14, while U-Net ranges from 3.25 K on Case 12 to 40.36 K on Case 13.
We next test whether this failure can be recovered with limited target-domain supervision. Using only labeled samples per OOD case, all eight models improve substantially, with MAE reductions of approximately 76–93%. The best MAE decreases to 1.71 K, and reaches 0.59 K at ; additional labels provide progressively smaller gains. Figure 3(b) additionally shows that worst-pixel errors remain several times larger than the global-field errors after 10-shot adaptation: even Therm-FM-T, which attains the lowest RMSE (2.09 K), retains a 6.70 K MaxAE.
Accuracy–Cost Trade-offs
Predictive accuracy also comes with substantially different model sizes and training costs. Table 6 compares S4 RMSE and MAE with parameter count and aggregate GPU time. Therm-FM-L achieves the strongest S4 within-family accuracy under both field metrics, at the cost of 628.4M parameters; U-FNO is the strongest operator baseline and stays within 1.16× of the best RMSE with only about 5.1M parameters. DeepOHeat is the cheapest model to train but has the highest S4 RMSE and MAE. Increasing model size yields only diminishing accuracy gains: scaling Therm-FM from B to L quadruples the parameters and increases training time by about 1.7×, yet reduces S4 RMSE by only about 4%.
| Method | RMSE | MAE | Params | GPU·s |
| FNO | 1.291 | 0.994 | 12.0M | 1,867 |
| U-Net | 1.271 | 0.976 | 5.06M | 875 |
| U-FNO | 0.939 | 0.724 | 5.10M | 11,658 |
| SAU-FNO | 0.953 | 0.739 | 5.11M | 22,300 |
| DeepOHeat | 4.739 | 3.903 | 7.92M | 264 |
| Therm-FM-T (21M) | 1.050 | 0.794 | 20.7M | 3,297 |
| Therm-FM-B (158M) | 0.846 | 0.618 | 157.6M | 6,580 |
| Therm-FM-L (629M) | 0.811 | 0.585 | 628.4M | 11,030 |
Table 6. S4 accuracy–cost comparison. RMSE and MAE are in kelvin; GPU·s is aggregate GPU time (wall time multiplied by GPU count). Lower is better for RMSE, MAE, and GPU·s.
These results further separate model capacity from predictive quality. U-FNO and SAU-FNO stay within 1.16× and 1.18× of the best S4 RMSE, respectively, at roughly 120× fewer parameters than Therm-FM-L, and scaling Therm-FM from 157.6M to 628.4M parameters yields only a marginal accuracy improvement. Together with the large variation in aggregate GPU time, this shows that S4 accuracy, model size, cross-package robustness, and computational efficiency are distinct model properties. No single architecture dominates all dimensions, motivating their separate reporting within the benchmark.
Key Findings and Implications
Finding 1: Broader represented thermal variability remains learnable. Across S2–S4, prediction errors increase gradually as represented variation in layout, material properties, and boundary conditions broadens, without the abrupt degradation observed under cross-package transfer. Implication: Within-Family Generalization should be evaluated across multiple ranges of physical variation rather than summarized by a single score.
Finding 2: Cross-package transfer challenges even the leading models. All evaluated models exhibit sharply higher global-field errors from S4 to S5. Therm-FM retains the best aggregate accuracy, but its zero-shot errors remain far above within-family levels and vary across target packages. Implication: Cross-Package OOD should remain a separate evaluation track, with both aggregate and per-case results.
| Method | ||||
| U-FNO | 27.11 | 2.23 | 1.06 | 0.59 |
| SAU-FNO | 25.17 | 2.16 | 1.08 | 0.71 |
| FNO | 39.47 | 2.67 | 1.17 | 0.72 |
| U-Net | 17.62 | 1.80 | 1.01 | 0.73 |
| DeepOHeat | 23.15 | 5.48 | 3.06 | 2.13 |
| Therm-FM-T (21M) | 8.04 | 1.71 | 1.04 | 0.72 |
| Therm-FM-B (158M) | 12.42 | 2.10 | 1.24 | 0.62 |
| Therm-FM-L (629M) | 14.66 | 2.48 | 1.02 | 0.67 |
Table 5. Target-domain adaptation on S5. Values are MAE in kelvin; denotes labeled samples per OOD case. uses the common adaptation holdout and is therefore slightly different from the full-set zero-shot score in Table 4.
Finding 3: Therm-FM leads most Scopes, but its advantage and scaling gains are not uniform. The Therm-FM family achieves the best global-field accuracy on S1, S2, S4, and S5, with a particularly pronounced advantage on S1, whereas U-FNO is the most accurate model on all four reported metrics at S3. The best-performing Therm-FM variant also changes with the Scope (L at S2 and S4, B at S3, T at S5), and larger models do not consistently improve accuracy. Implication: comparisons should consider Scope-specific absolute errors, relative improvements, and scaling behavior, rather than rankings or parameter counts alone.
Finding 4: Adaptation is effective but distinct from zero-shot generalization. On the common holdout, limited target-domain supervision recovers a substantial fraction of the zero-shot accuracy loss, but adapted performance no longer measures frozen cross-package transfer. Implication: zero-shot generalization and target-domain sample efficiency should be evaluated separately, with the label budget per target case reported explicitly.
Discussion and Limitations
Benchmark scope. IC-ThermBench focuses on generalization in learning-based thermal prediction. S2–S5 cover HotSpot-generated steady-state 2.5D chiplet systems, while S1 provides fixed-design steady-state, transient, and industrial-package reference tasks. Current results do not characterize all industrial package settings. Future extensions may include multi-tier 3D structures, alternative cooling technologies, transient generated tasks, and higher-fidelity data.
Data coverage density. S2–S4 broaden physical support with comparable sample counts. Their increasing errors therefore reflect the combined difficulty of broader distributions rather than the isolated effect of any added variable. Sparser coverage may also contribute to the degradation. Larger or more targeted datasets could help distinguish model limitations from insufficient data coverage.
Pretraining transfer under structural and physical variation. Therm-FM’s generally smaller relative margins on S2–S5 may reflect the shift from fixed-design prediction to broader structural and physical variation. Within each S1 task, geometry, materials, and cooling conditions remain fixed while power inputs and, for transient tasks, time vary. This fixed-domain source-to-solution setting may favor reuse of PDE-pretrained representations. In contrast, S2–S4 introduce layout, material, and boundary-condition variation, while S5 requires transfer to unseen packages. These demands may limit direct pretraining transfer, reducing Therm-FM’s relative advantage and, at S3, eliminating its lead.
This explanation remains a hypothesis: the source-suite and generated tasks also differ in task definitions and fidelity, so the observed change cannot be attributed to structural variation alone. Geometry-aware representations and pretraining over variable structures and boundary conditions merit further study.
Simulator-domain validity. Because S2–S5 use HotSpot labels, the results measure prediction and generalization within that simulation domain rather than silicon-level or signoff-level accuracy. Deployment still requires validation against higher-fidelity simulation or measurements.
Conclusion
We presented IC-ThermBench, an open benchmark for evaluating generalizable 2.5D/3D-IC thermal learning under a common evaluation contract. Five Generalization Scopes connect established 3D-IC fixed-design tasks with progressively broader layout, material, and boundary-condition variation, followed by Cross-Package OOD evaluation. Across eight representative baselines, observed physical variability remains largely learnable, whereas transfer to unseen package systems exposes a much larger generalization gap; limited target-domain supervision can recover substantial accuracy but does not replace zero-shot generalization. By providing shared data, fixed splits, input semantics, metrics, generation tools, and reproducible baselines, IC-ThermBench establishes a common foundation for developing and comparing thermal predictors beyond narrowly defined design distributions.
References
Benchmark Reporting and Fair-Comparison Contract
IC-ThermBench results are directly comparable on the primary benchmark only when they use the same dataset version, immutable splits, input and label definitions, preprocessing, normalization, and reference metrics. Training hardware, parameter count, optimization configuration, and result provenance must also be reported. Frozen zero-shot evaluation and target-domain adaptation are separate tracks, and any alternative simulator, resolution, split, preprocessing, or metric must be reported as a separate track. Table 7 summarizes the minimum reporting record.
| Item | Required record |
| Data | Version, case list, split hash |
| Labels | Simulator, version, grid, tolerances, boundary model |
| Inputs | Channel semantics, normalization, resolution handling |
| Zero-shot | Frozen checkpoint and unchanged preprocessing |
| Adaptation | per case, pool/holdout split, nested indices |
| Metrics | RMSE, MAE, , MaxAE, , Top-50 MAE |
| Optimization | Loss, learning rate, epochs, checkpoint rule |
| Compute | Device, GPU count, wall time, GPU·s, parameters |
| Provenance | Author-reported, released-code, or reimplementation |
Table 7. Minimum record for a comparable benchmark result.
Physical Formulation and Dataset Specification
Thermal Formulation
Let denote the package domain, its temperature, the mass density, the specific heat, the thermal conductivity, and the volumetric heat source. The transient field obeys
Steady-state analysis removes the temporal term. On the boundary, the reference problem may impose
for Dirichlet, prescribed-flux, and convective conditions. Across a material interface , temperature and normal heat flux remain continuous,
Transient tasks additionally specify . For a fixed structural configuration , these equations define a solution operator from power, material, boundary, and temporal variables to the temperature field. S1 preserves both steady and transient source tasks; generated S2–S5 tasks are steady state.
HotSpot Generation Configuration
All generated S2–S5 labels use the open-source HotSpot grid-based detailed-3D flow at 64 × 64 resolution. The exact executable, configuration files, layer files, units, and conversion scripts are released with the benchmark artifact. Temperatures are exported in kelvin. The six layer stack is fixed across S2–S5; only the benchmark variables listed below change.
Parameter Distributions
For S3 and above, filler conductivity is sampled log-uniformly from [0.2, 5.0] W/(m·K), while silicon remains fixed at 100 W/(m·K). For S4–S5, ambient temperature is sampled uniformly from [25, 55] °C, and the convection coefficient is sampled log-uniformly from [500, 20000] W/(m²·K). Convective resistance is derived from and heatsink geometry rather than sampled independently.
Input Tensor Semantics
The cumulative model inputs are
The power channel is each chiplet’s nominal power in watts broadcast over the cells occupied by that chiplet; it is not pixel-wise power density. Coordinates are physical millimeters. Conductivity is a cell-wise material field, while , , and are sample-global scalars broadcast over the grid. Each input channel is normalized independently using statistics fitted on the training subset only.
Converted tensors use float32, a 64 × 64 grid, and a retained singleton depth . The released S2Vs additionally retain chiplet identifiers, occupancy and edge masks, and normalized coordinates for reconstruction, but these auxiliary columns are not given to the evaluated models.
Layout Generation
Randomized layouts use random, dispersed, cluster, and boundary modes in a 6:2:1:1 ratio. Layout generation enforces package boundaries and non-overlap, varies legal position and orientation, and preserves each chiplet’s case-level nominal power and dimensions. Accepted chiplets are rasterized onto the 64 × 64 grid; occupied cells inherit chiplet identity and nominal power, while background cells remain zero. The released case manifest and generator configuration define the exact legal-layout constraints.
Independence of S2–S4
S2–S4 use the same Cases 1–10 but are generated and sampled independently. They are not paired samples to which channels are progressively added. Their comparison therefore measures progressively broader represented distributions, not a sample-wise one-factor ablation.
| # | Layer (bottom to top) | Power | Heat capacity | Resistivity | Conductivity | Thickness |
| 0 | Substrate | No | 3.33 | |||
| 1 | Epoxy SiO underfill + C4 | No | 0.625 | |||
| 2 | Silicon interposer | No | 0.01 | 100 | ||
| 3 | Underfill with microbumps | No | 0.625 | |||
| 4 | Chip silicon / filler | Yes | / variable | / – | ||
| 5 | TIM | No | 0.25 | 4.0 |
Table 8. Six-layer HotSpot stack used for generated S2–S5 tasks.
| Channel | S2 | S3 | S4 | S5 OOD |
| Nominal chiplet power (W) | 0–300 | 0–300 | 0–300 | 0–280 |
| Physical (mm) | 0–59 | 0–59 | 0–59 | 0–72.7 |
| Physical (mm) | 0–61 | 0–61 | 0–61 | 0–72.7 |
| (W/(m·K)) | – | 0.2–100 | 0.2–100 | 0.2–100 |
| (K) | – | – | 298.2–328.1 | 298.2–328.1 |
| (W/(mK)) | – | – | 500–19,990 | 500–20,000 |
| (K/W) | – | – | 0.00087–0.2081 | 0.00059–0.1080 |
Table 9. Empirical model-input ranges. Dashes denote absent channels. For the cell-wise conductivity channel , the upper value 100 W/(m·K) comes from fixed silicon cells; only filler conductivity is sampled, over W/(m·K).
| # | Channel | Unit | Semantics | |||
| 0 | chiplet_power | W | Yes | Yes | Yes | Nominal chiplet power broadcast over occupied cells; zero in background |
| 1 | grid_x | mm | Yes | Yes | Yes | Physical coordinate |
| 2 | grid_y | mm | Yes | Yes | Yes | Physical coordinate |
| 3 | local_thermal_k | W/(m·K) | – | Yes | Yes | Cell-wise material conductivity |
| 4 | ambient_K | K | – | – | Yes | Sample-global ambient temperature |
| 5 | h_w_m2k | W/(mK) | – | – | Yes | Sample-global convection coefficient |
| 6 | r_convec_k_per_w | K/W | – | – | Yes | Derived convective resistance |
Table 10. Model-input channels for generated S2–S5 tasks.
| Subset | Per case | Per Scope |
| Training | 1,080 | 10,800 |
| Validation | 120 | 1,200 |
| Test | 300 | 3,000 |
Table 11. Case-balanced split for each of S2–S4.
Splits, OOD Protocol, and Metrics
Case-Balanced S2–S4 Splits
Samples are interleaved round-robin by case. Within each case, 1,080 samples are used for training, 120 for validation, and 300 for testing, giving 10,800/1,200/3,000 samples per Scope. Normalization statistics are fitted only on the 10,800-sample training subset and remain fixed for validation and test.
Official S5 Zero-Shot Protocol
Cases 11–15 are excluded from source training, validation, normalization fitting, and checkpoint selection. The official zero-shot track evaluates the frozen S4 checkpoint on all 5,000 S5 samples with unchanged preprocessing and no target labels. Because an unseen package can shift both structure and the empirical support of model-visible coordinates or derived variables, this track is termed Cross-Package OOD, not a controlled structural-only intervention.
Target-Domain Adaptation Protocol
For each held-out case, the first 500 released samples form an adaptation pool and the final 500 form a fixed evaluation holdout. The released nested index lists define labels per case, so the total target-label budget is . Each run starts from the frozen S4 checkpoint, reinitializes the optimizer, retains the source normalization, and fine-tunes for 50 epochs. Non-Therm-FM models use Adam with learning rate , weight decay , and batch size 20; Therm-FM uses , one tenth of its training rate. The reported tables use the epoch-50 checkpoint.
| Case | Principal stress | Difference from Cases 1–10 |
| 11 | Chiplet count | 80 chiplets versus a training maximum of 61; utilization is reduced to 32% for feasible generation |
| 12 | Power density | 0.85 W/mm versus a training maximum of 0.745 W/mm |
| 13 | Size heterogeneity | One 42 × 16 mm die plus 24 micro-chiplets of 2–2.6 mm |
| 14 | Power concentration | One chiplet contributes 90.9% of total power versus a training maximum of 69.2% |
| 15 | Package utilization | 68.1% utilization versus a training maximum of 67.1% |
Table 12. Principal stress characteristics of held-out S5 systems. Each case primarily stresses the listed attribute but is not a strict single-factor intervention.
| Model | Configuration | |||
| U-FNO | Net3d; modes (10, 10, 1); width 36 | 5,103,905 | 5,103,941 | 5,104,049 |
| SAU-FNO | SAUNet3d; modes (10, 10, 1); width 36 | 5,107,901 | 5,107,937 | 5,108,045 |
| FNO | FNO3d; modes (12, 12, 1); width 72 | 11,974,721 | 11,974,793 | 11,975,009 |
| U-Net | Channels (58, 116, 232, 348); three max-pools; nearest-neighbor upsampling; skip connections; no BN/dropout | 5,060,037 | 5,060,559 | 5,062,125 |
| DeepOHeat | Frozen 128-D Gaussian Fourier coordinate trunk; 128-wide 3-layer SiLU trunk; flattened-input 256-wide 7-layer branch; inner product + bias | 3,721,985 | 4,770,561 | 7,916,289 |
| Therm-FM T/B/L | scOT/Poseidon pretrained checkpoints | 21M / 158M / 629M (paper-reported) |
Table 13. Model configurations and input-dependent parameter counts for S2–S4 ().
The adaptation curve’s value is the frozen checkpoint evaluated on the common 2,500-sample holdout. It therefore differs slightly from the official zero-shot score, which uses all 5,000 S5 samples.
Metric Definitions
For sample , let at grid location . The per-sample field metrics are
Top-50 MAE is the mean absolute error over the 50 hottest locations selected from the ground-truth field. The predicted and true maxima need not occur at the same location for . MAE, RMSE, MaxAE, , and Top-50 MAE are computed per sample and then averaged arithmetically over the test samples; the reported MaxAE is therefore the mean per-sample worst-pixel error rather than the single worst pixel of the whole test set. For multi-layer fields (S1), each metric is first averaged over the layers of a sample. All eight baselines share one metric implementation. The benchmark pools all samples and grid locations,
Baseline and Training Details
Baseline Architectures
Because , the third Fourier mode count is one. Increasing the channel count changes only the input layer for convolutional and Fourier baselines, but substantially enlarges DeepOHeat’s flattened branch.
Optimization Configuration
All evaluated methods are label-supervised. For DeepOHeat this differs from the original residual-only training formulation and keeps the benchmark question fixed: given identical simulated labels, which model best predicts the temperature field? FNO, U-FNO, SAU-FNO, U-Net, and DeepOHeat use the final-epoch checkpoint with early stopping disabled; Therm-FM retains the validation-best checkpoint selected by its scOT training pipeline.
Software and Hardware
Experiments use NVIDIA RTX A6000 48GB GPUs and an AMD EPYC 7543 CPU with PyTorch 2.0.1 and CUDA 11.8. All baselines use a single GPU. Wall time covers model training only and excludes one-time HotSpot generation, format conversion, checkpoint download, and manual hyperparameter exploration.
Executable Interface and Model Extension
The anonymized benchmark artifact accompanying this submission exposes one entry point for training. Within-Family Generalization testing, frozen Cross-Package OOD evaluation, and target-case adaptation.
| Setting | FNO/U-FNO/SAU-FNO/U-Net/DeepOHeat | Therm-FM T/B/L |
| Optimizer | Adam | Adam |
| Learning Rate | (embedding: ) | |
| Weight decay | ||
| Schedule | StepLR(step 2, ) | Cosine; zero warmup |
| Batch size | 20 | 40 |
| Epochs | 100 | 100 |
| Objective | Normalized-space MSE | scOT configured MSE; max grad norm 5 |
| Checkpoint | Final epoch | Validation-best |
| Hardware | One RTX A6000 48GB | One RTX A6000 48GB |
Table 14. Training configuration. All normalizers are fitted on training data only and applied per channel.
Every method is adapted to the same repository-level tensor contract: each sample pairs a -channel input field with a temperature field on the same grid, where the current generated release uses and . A conventional PyTorch model needs only an implementation and a registry entry that declares its recipe and channel-dependent constructor; the shared runner then supplies the immutable split, normalization, six metrics, JSON output, and S5 adaptation loop. Adding a conventional PyTorch baseline requires one implementation file and one short MODEL_ZOO entry containing its metric prefix, optimizer recipe, adaptation rate, and a constructor that receives , , and grid size. Models with custom training stacks use the same tensor and metric contracts through a separate adapter.
This interface separates architecture-specific tensor conversion and optimization from benchmark-controlled data and evaluation. A released result can therefore be regenerated with the same sample identifiers and metric implementation, while a new model can be compared without copying or silently modifying the evaluation code.
S1 Source-Suite Results
S1 comprises eleven fixed-design source tasks, including steady-state and transient Alpha EV6 configurations and two industrial packages. Table 3 summarizes the S1 results: the Therm-FM values coincide with those reported in the Therm-FM study [9], whereas the other baselines are trained and evaluated in the current benchmark. Per-task tables are omitted for conciseness.
Qualitative observations. Figure 4(a) shows that the selected Therm-FM results preserve the global temperature distribution and dominant hotspot locations across the three HotSpot cases and two industrial packages; the visible residuals remain localized mainly near sharper thermal transitions. Panel (b) shows close agreement through the sampled transient steps for HS-SC and HS-OC, including the temperature rise and spatial diffusion from 0 to 8 ms. These observations are qualitative and refer only to the source Therm-FM visualizations; the corresponding quantitative results are reported in the Therm-FM study cited above.

Figure 4. Qualitative S1 visualizations reproduced directly from the best-performing Therm-FM source configurations [9]; they were not regenerated for this paper.
Complete Within-Family Generalization Results (S2–S4)
The main paper retains four metrics per Scope to expose the central trend: RMSE and MAE for global field accuracy, together with worst-pixel MaxAE and Top-50 MAE. Tables 15–17 restore all six reported metrics for every baseline. The datasets are independently generated under a common case-balanced protocol; the progression should therefore be read as increasing represented-distribution difficulty, not as a paired causal ablation on the same samples.
Reading S2. Therm-FM L leads all six metrics, with U-FNO and Therm-FM B close behind on the global field errors; among the operator baselines U-FNO leads SAU-FNO on every metric except MaxAE, and model scale helps Therm-FM monotonically here. DeepOHeat trails by a wide margin. U-Net ranks lower on the worst-pixel and hotspot-region errors (MaxAE and Top-50 MAE) than on global RMSE, which also illustrates why hotspot-region and field-wide errors should be reported together.
Reading S3: Material. U-FNO leads all six metrics once conductivity becomes an explicit input field, ahead of SAU-FNO and every Therm-FM variant. The best RMSE rises by about 89% relative to S2, yet the best remains above 0.998, so represented material variation increases difficulty without causing a generalization break. Model scale is still not a reliable predictor: Therm-FM L does not improve over Therm-FM B here.
Metric-wise behavior across represented support. The six metrics expose a stable distinction between reconstructing the full field and tracking the thermally critical region. SAU-FNO is never the best global predictor, yet its MaxAE is the second best on every within-family Scope; a deployment concerned mainly with bounding the worst-case local error could therefore reach a different model choice from one that requires an accurate temperature field everywhere. Conversely, DeepOHeat’s global and hotspot errors deteriorate together as input dimensionality grows, suggesting that its present flattened branch is a limiting factor under the representation rather than a metric-specific failure. U-FNO and SAU-FNO remain comparatively balanced across all six criteria.
Implication for progressive evaluation. No architecture should be described as generally superior from one Scope. The leading model changes with the Scope (Therm-FM L at S2 and S4, U-FNO at S3), the best Therm-FM variant switches from L to B at S3, and — as S5 shows next — the best within-family model is not the best cross-package one. This is informative: the configuration that best absorbs represented placement, material, and boundary variation need not be the one that transfers to unseen packages. The progression therefore functions as a capability profile rather than a single scalar leaderboard. The first Scope at which error rises sharply indicates where performance begins to degrade as broader physical variation is introduced.
Reading S4: Boundary. Therm-FM L leads all six metrics, with U-FNO the strongest operator baseline on the global-field errors and SAU-FNO on MaxAE. Relative to S2, the best RMSE increases by roughly , while the best peak-temperature error stays below 0.5 K. The distinction matters for deployment: a predictor may be sufficiently accurate for thermal-limit screening even when its full-field error has increased. Most importantly, this moderate within-family degradation is qualitatively smaller than the case-disjoint failure reported next.
| Method | RMSE | MAE | MaxAE | Top-50 MAE | ||
| FNO | 0.9619 | 0.7200 | 0.9967 | 4.4718 | 0.4826 | 0.5059 |
| U-Net | 0.8432 | 0.6302 | 0.9975 | 5.0203 | 0.4753 | 0.5017 |
| U-FNO | 0.6700 | 0.4928 | 0.9984 | 3.3669 | 0.3312 | 0.3730 |
| SAU-FNO | 0.7078 | 0.5255 | 0.9982 | 3.6997 | 0.4080 | 0.4466 |
| DeepOHeat | 3.9699 | 3.2259 | 0.9437 | 10.5532 | 3.5631 | 4.8260 |
| Therm-FM T | 0.9105 | 0.6787 | 0.9968 | 4.2679 | 0.4989 | 0.5507 |
| Therm-FM B | 0.7215 | 0.5165 | 0.9980 | 3.7171 | 0.3378 | 0.3781 |
| Therm-FM L | 0.7366 | 0.5268 | 0.9979 | 3.7848 | 0.3436 | 0.3904 |
Table 16. Complete S3 Material results under Within-Family Generalization.
Complete Cross-Package OOD and Adaptation Results
This section reports the complete S5 results behind the condensed main-paper discussion. The official zero-shot test takes each baseline’s frozen S4 checkpoint, trained under the full seven-channel material-and-boundary protocol, and evaluates it directly on all 5,000 samples from held-out cases C11–C15, without using those cases for training, validation, normalization, or checkpoint selection. No S5-specific training is performed; the model, its input schema, and its normalization statistics are exactly those of S4. It is therefore a Cross-Package OOD test rather than an isolated single-parameter extrapolation.
The aggregate and per-case views answer complementary questions. The aggregate table measures expected performance over the released stress set and supports a compact leaderboard. The per-case table reveals whether that average reflects consistent transfer or success on one system offsetting failure on another. We report both because the five held-out systems stress different combinations of chiplet count, die outline, size mixture, power density, power concentration, and utilization; they are not repeated samples from one homogeneous OOD population.
The official zero-shot record also remains separate from the adaptation baseline below. The former evaluates all 5,000 released S5 samples; the latter evaluates the common 2,500-sample holdout reserved for comparing label budgets. Mixing these denominators would make an apparent adaptation gain partly a test-set change. Keeping both records visible makes the zero-shot-to-few-shot transition auditable.
No single baseline dominates every held-out package. Therm-FM T wins the pooled six-metric comparison and is the best model on C14 and C15, but its advantage comes mainly from comparatively balanced errors (6.3–11.2 K on every case) rather than from winning every case: Therm-FM B is best on C11 and C13 and U-Net on C12, showing that the aggregate OOD gap combines several distinct package-level shifts. DeepOHeat is also strong on the high-chiplet-count C11 case (4.79 K) but degrades on C12, C14, and C15; SAU-FNO and U-Net show the opposite kind of specialization, with their lowest errors on the high-power-density C12 case. Moreover, larger Therm-FM variants transfer progressively worse on five of the six metrics (all but Top-50 MAE): zero-shot RMSE rises from 8.67 K (T) to 13.06 K (B) and 15.34 K (L). These ranking reversals show that strong within-family accuracy under represented physical variation does not imply robust cross-package generalization.
Target-Case Adaptation
For each held-out case, adaptation uses the released fixed nested index sets for samples from a 500-sample adaptation pool (identical for every model) and evaluates on the separate 500-sample holdout. Counts are per case, so uses 50 labels in total. The column below evaluates the frozen checkpoint on the same common 2,500-sample holdout; it is not a replacement for the official 5,000-sample zero-shot result in Table 18.
| Method | RMSE | MAE | MaxAE | Top-50 MAE | ||
| FNO | 42.1755 | 39.3575 | -1.6429 | 65.9182 | 59.9215 | 39.7131 |
| U-Net | 18.8542 | 17.3739 | 0.3776 | 32.4320 | 24.0189 | 18.0644 |
| U-FNO | 28.1538 | 27.0604 | -0.2399 | 39.4894 | 31.6274 | 26.6474 |
| SAU-FNO | 26.1208 | 25.0555 | -0.1239 | 36.3580 | 31.7934 | 26.9022 |
| DeepOHeat | 23.9735 | 23.1409 | 0.0759 | 33.4330 | 25.6216 | 24.0784 |
| Therm-FM T | 8.6729 | 8.0029 | 0.9101 | 15.5347 | 10.4522 | 9.0454 |
| Therm-FM B | 13.0554 | 12.3390 | 0.6921 | 21.2864 | 16.7420 | 15.7459 |
| Therm-FM L | 15.3361 | 14.6325 | 0.6835 | 24.0876 | 17.1175 | 13.7678 |
Table 18. Complete zero-shot results on the official S5 test set. All error metrics are in kelvin; higher is better and lower is better for the other metrics.
Controlled adaptation comparison. Every run starts from its released S4 checkpoint, retains the source normalizers, reinitializes only the optimizer, and trains for 50 epochs. The target holdout is never used to fit normalization or update weights. The non-Therm-FM models use their registered adaptation rate of and batch size 20; Therm-FM uses , one tenth of its training rate, because the larger rate used by the operator baselines is unstable for its pretrained stack. These choices preserve architecture-appropriate optimization while holding the data, labels, split membership, metric code, and final-checkpoint rule fixed.
Reading the label budget. is defined per held-out system rather than as one pooled count. Thus means 50 labels in total and guarantees that even the smallest budget observes every target case. This protocol tests low-cost target supervision, not zero-shot generalization, and the two results must not be conflated. A method that improves the frozen S5 score addresses cross-package transfer, whereas one that reaches a lower adapted score with the same addresses target-domain sample efficiency. The complete curve is required because those rankings need not agree.
The complete tables reinforce two points hidden by RMSE alone. First, the best model changes with both label budget and metric: Therm-FM T leads all six metrics at and all but Top-50 MAE at , where U-Net leads; at the lead is split between Therm-FM L (RMSE and MaxAE) and U-Net (MAE, , and Top-50 MAE), with the two tied on ; at Therm-FM L co-leads RMSE with U-FNO and leads every other metric except Top-50 MAE (U-Net); and at the operator baselines take over the global-field metrics (SAU-FNO at , tied with Therm-FM L on ; U-FNO at ) and SAU-FNO shows the worst-pixel error, while the hotspot-oriented metrics are led by Therm-FM B at and Therm-FM L at . Second, the first 10 labels per case recover most of the global correlation ( for every model), but worst-pixel errors remain several kelvin and improve more slowly.
| Method | ||||||
| U-FNO | 28.19 | 2.65 | 1.29 | 1.06 | 1.10 | 0.76 |
| SAU-FNO | 26.23 | 2.56 | 1.32 | 1.20 | 0.86 | 0.90 |
| FNO | 42.27 | 3.30 | 1.47 | 1.24 | 1.01 | 0.92 |
| U-Net | 19.10 | 2.26 | 1.29 | 1.21 | 0.95 | 0.93 |
| DeepOHeat | 23.98 | 6.22 | 3.69 | 3.28 | 2.91 | 2.71 |
| Therm-FM T | 8.71 | 2.09 | 1.32 | 1.10 | 0.92 | 0.92 |
| Therm-FM B | 13.13 | 2.45 | 1.48 | 1.18 | 0.99 | 0.81 |
| Therm-FM L | 15.36 | 2.88 | 1.27 | 1.06 | 0.92 | 0.86 |
Table 20. RMSE (K) after target-case adaptation. Each is the number of labeled adaptation samples per held-out case.
| Method | 0 | 10 | 50 | 100 | 250 | 500 |
| U-FNO | 27.11 | 2.23 | 1.06 | 0.86 | 0.92 | 0.59 |
| SAU-FNO | 25.17 | 2.16 | 1.08 | 0.98 | 0.69 | 0.71 |
| FNO | 39.47 | 2.67 | 1.17 | 0.98 | 0.80 | 0.72 |
| U-Net | 17.62 | 1.80 | 1.01 | 0.97 | 0.75 | 0.73 |
| DeepOHeat | 23.15 | 5.48 | 3.06 | 2.67 | 2.34 | 2.13 |
| Therm-FM T | 8.04 | 1.71 | 1.04 | 0.87 | 0.71 | 0.72 |
| Therm-FM B | 12.42 | 2.10 | 1.24 | 0.96 | 0.81 | 0.62 |
| Therm-FM L | 14.66 | 2.48 | 1.02 | 0.85 | 0.72 | 0.67 |
Table 21. Complete target-case adaptation results for (a) MAE (K) and (b) pooled . Each is the number of labels per held-out case. (a) MAE (K); lower is better.
| Method | 0 | 10 | 50 | 100 | 250 | 500 |
| U-FNO | -0.2957 | 0.9889 | 0.9977 | 0.9985 | 0.9985 | 0.9993 |
| SAU-FNO | -0.1794 | 0.9897 | 0.9974 | 0.9980 | 0.9990 | 0.9990 |
| FNO | -1.7581 | 0.9875 | 0.9976 | 0.9983 | 0.9988 | 0.9990 |
| U-Net | 0.3336 | 0.9938 | 0.9980 | 0.9983 | 0.9989 | 0.9990 |
| DeepOHeat | 0.0361 | 0.9256 | 0.9784 | 0.9842 | 0.9883 | 0.9899 |
| Therm-FM T | 0.9056 | 0.9946 | 0.9979 | 0.9985 | 0.9989 | 0.9989 |
| Therm-FM B | 0.6738 | 0.9923 | 0.9971 | 0.9982 | 0.9987 | 0.9992 |
| Therm-FM L | 0.6698 | 0.9889 | 0.9980 | 0.9986 | 0.9990 | 0.9990 |
Table 21. (b) Pooled ; higher is better.
| Method | 0 | 10 | 50 | 100 | 250 | 500 |
| U-FNO | 39.50 | 8.68 | 5.94 | 5.21 | 4.92 | 4.47 |
| SAU-FNO | 36.49 | 7.87 | 4.86 | 4.55 | 3.71 | 3.66 |
| FNO | 66.03 | 10.91 | 5.59 | 4.71 | 4.02 | 3.82 |
| U-Net | 32.68 | 9.42 | 7.40 | 6.85 | 5.97 | 5.54 |
| DeepOHeat | 33.49 | 12.57 | 9.34 | 8.82 | 8.21 | 8.12 |
| Therm-FM T | 15.60 | 6.70 | 5.28 | 4.84 | 4.45 | 4.49 |
| Therm-FM B | 21.34 | 7.08 | 5.15 | 4.66 | 4.34 | 4.45 |
| Therm-FM L | 24.16 | 7.96 | 4.81 | 4.39 | 4.24 | 4.24 |
Table 22. Complete target-case adaptation results for (a) per-sample MaxAE and (b) peak-temperature error , both in kelvin. (a) MaxAE; lower is better.
| Method | 0 | 10 | 50 | 100 | 250 | 500 |
| U-FNO | 31.68 | 3.07 | 1.33 | 1.01 | 0.94 | 0.82 |
| SAU-FNO | 31.94 | 3.16 | 1.72 | 1.29 | 0.86 | 1.16 |
| FNO | 60.10 | 3.03 | 1.47 | 1.35 | 0.87 | 0.88 |
| U-Net | 24.41 | 2.48 | 1.23 | 1.06 | 0.93 | 0.92 |
| DeepOHeat | 25.62 | 6.78 | 4.12 | 3.90 | 3.48 | 3.28 |
| Therm-FM T | 10.53 | 2.39 | 1.55 | 1.24 | 0.92 | 0.99 |
| Therm-FM B | 16.82 | 2.90 | 1.45 | 1.09 | 0.77 | 0.77 |
| Therm-FM L | 17.21 | 3.21 | 1.26 | 0.93 | 0.83 | 0.65 |
Table 22. (b) ; lower is better.
| Method | ||||||
| U-FNO | 26.65 | 3.15 | 1.49 | 1.12 | 0.99 | 0.88 |
| SAU-FNO | 26.99 | 3.32 | 1.89 | 1.45 | 0.95 | 1.20 |
| FNO | 39.86 | 3.30 | 1.75 | 1.49 | 0.98 | 1.02 |
| U-Net | 18.39 | 2.24 | 1.28 | 1.02 | 1.00 | 0.91 |
| DeepOHeat | 24.08 | 6.76 | 4.61 | 4.21 | 3.75 | 3.59 |
| Therm-FM T | 9.08 | 2.37 | 1.64 | 1.37 | 1.10 | 1.08 |
| Therm-FM B | 15.77 | 3.05 | 1.71 | 1.40 | 0.91 | 0.95 |
| Therm-FM L | 13.78 | 2.96 | 1.45 | 1.04 | 0.98 | 0.86 |
Table 23. Complete Top-50 MAE (K) under target-case adaptation. The metric is computed at the 50 hottest ground-truth locations of each sample.
Figure 5. Target-case adaptation on S5 for all six benchmark metrics. is the number of labeled samples per held-out case ( is the frozen S4 checkpoint on the common holdout); error metrics use a logarithmic axis. The steep improvement at the first few labeled samples is consistent across model families, while the remaining gaps depend on the error criterion.
Optimization record. All adaptation tables use the final checkpoint after the same 50-epoch budget and evaluate every five epochs. The release stores the intermediate JSON records so that convergence and checkpoint-selection studies can be reconstructed without changing the official final-checkpoint comparison. The appendix freezes the complete final-checkpoint numbers for the eight registered baselines; intermediate traces remain machine-readable rather than being summarized by an additional, potentially redundant figure.
Training Cost and Efficiency Details
Table 24 reports model-training cost under the same released splits and hardware protocol. Wall time measures the standard training run, while GPU-s multiplies wall time by the number of GPUs and is included only to expose aggregate accelerator occupancy. Neither quantity includes one-time HotSpot label generation, data conversion, checkpoint download, or manual hyperparameter exploration.
| Method | Params. | S2 Wall (s) | S2 GPU-s | S3 Wall (s) | S3 GPU-s | S4 Wall (s) | S4 GPU-s |
| FNO | 11.98M | 1844 | 1844 | 1857 | 1857 | 1867 | 1867 |
| U-Net | 5.06M | 845 | 845 | 841 | 841 | 875 | 875 |
| U-FNO | 5.10M | 11603 | 11603 | 11606 | 11606 | 11658 | 11658 |
| SAU-FNO | 5.11M | 22193 | 22193 | 22299 | 22299 | 22300 | 22300 |
| DeepOHeat | 7.92M | 194 | 194 | 228 | 228 | 264 | 264 |
| Therm-FM T | 20.75M | 3195 | 3195 | 3245 | 3245 | 3297 | 3297 |
| Therm-FM B | 157.63M | 6283 | 6283 | 6330 | 6330 | 6580 | 6580 |
| Therm-FM L | 628.38M | 10782 | 10782 | 10806 | 10806 | 11030 | 11030 |
Table 24. Parameter count and training cost across S2–S4. GPU-s is wall time multiplied by GPU count and should not be read as hardware-normalized throughput. All methods use a single GPU, so wall time and GPU-s coincide.
The table is intended for reproducibility, not for ranking architecture efficiency across different accelerators. All baselines run on a single GPU, so wall time and aggregate GPU-s coincide here.
DeepOHeat has the shortest wall time but also the largest errors across the S2–S4 within-family evaluations, so speed alone is not a sufficient selection criterion. Among the strong Fourier baselines, U-FNO is substantially cheaper than SAU-FNO under the recorded recipe, and it is also more accurate on the global-field errors of every within-family Scope; SAU-FNO’s additional attention cost mainly buys a lower worst-pixel error at S2 and S4. Therm-FM T has a modest wall time; the B and L variants increase aggregate accelerator use by about 2.0× and 3.3× at S4, improve within-family accuracy over T on every within-family Scope (although L does not improve over B at S3), and transfer worse in frozen OOD evaluation. These comparisons motivate reporting error, wall time, GPU count, and parameter scale together.
Cost-accounting boundary. The measured interval begins with the first optimizer update and ends after the final training epoch. It excludes one-time case generation, HotSpot label production, CSV-to-tensor conversion, checkpoint download, and manual hyperparameter exploration. Label generation is excluded because it is a dataset-production cost shared by all fully supervised models; for target adaptation it becomes a deployment-specific cost and should be reported separately. This boundary makes the model-training numbers reproducible without presenting them as end-to-end system latency.
Why wall time and GPU-s are both retained. Wall time approximates machine turnaround on the recorded machine, whereas GPU-s exposes aggregate occupancy. Neither is hardware-normalized throughput: four-GPU execution can reduce elapsed time while increasing total accelerator use, and different kernels exploit the RTX A6000 differently. A new result should therefore state device model, GPU count, precision, software stack, epoch count, batch size, and parameter count alongside both measures. When hardware changes, the released commands and output schema remain comparable even if raw timing does not.
Accuracy–cost selection. The table supports deployment choices rather than one universal efficiency rank. A fast screening loop may prefer U-Net despite lower accuracy; a Within-Family Generalization study may justify the additional cost of U-FNO or SAU-FNO; and a pretrained Therm-FM checkpoint may be attractive when S1 accuracy is central. Cross-Package OOD remains the decisive caveat: none of the cost advantages makes the frozen S5 errors operationally small, so adaptation label and optimization costs must be included whenever few-shot recovery is part of the intended workflow.
Generated-Scope Visualizations
Figure 6–Figure 10 show representative predictions for S2–S5 using the models specified in the corresponding captions, in the same format as the S1 visualizations (Figure 4). For S2–S4, columns correspond to Cases 1, 3, 5, 7, and 9 from each Scope’s own test split. For S5, columns correspond to Cases 11–15 under zero-shot evaluation and after target-case adaptation using labeled samples per case. Each column shows the median-RMSE sample within that case for the displayed model and evaluation setting. Ground truth and prediction share the temperature scale above each column, with out-of-range predictions clipped for visualization; the absolute-error panel uses a separate scale below.
Figure 6. S2 visualization with U-FNO. Columns correspond to Cases 1, 3, 5, 7, and 9, with the median-RMSE test sample shown for each case. Rows show ground truth, prediction, and absolute error in kelvin.
Figure 10. S5 visualization with U-FNO after target-case adaptation using labeled samples per case. Columns correspond to Cases 11–15, with the median-RMSE sample of each case shown from the common 2,500-sample holdout. Rows show ground truth, prediction, and absolute error in kelvin. The adapted predictions recover the global temperature structure on all five cases, including the scattered micro-chiplet hot spots of Case 13.
Reading S2–S4. Within the represented system families the predicted fields are visually indistinguishable from the HotSpot ground truth at the scale of the figure. The residual error is not spatially uniform: it concentrates along chiplet boundaries, where the temperature gradient is steepest, and at the corners of high-power dies, while the package interior and the background are reproduced to a fraction of a kelvin. Adding the material channel (S3) and the boundary channels (S4) raises the error scale at the worst pixel from roughly 1.7–2.9 K (S2) to 1.9–4.3 K (S3) and 6.1–9.1 K (S4); at S4 this worst pixel lies at a grid corner in every displayed sample. The error pattern keeps the same edge-localized character, and the additional physics is absorbed rather than breaking the predictor. These are the same median samples whose Scope-level averages appear in Table 15–Table 17.
| Method | RMSE | MAE | MaxAE | Top-50 MAE | ||
| FNO | 0.8284 | 0.6207 | 0.9977 | 3.9652 | 0.4667 | 0.4943 |
| U-Net | 0.7375 | 0.5412 | 0.9983 | 5.2731 | 0.4492 | 0.4996 |
| U-FNO | 0.4442 | 0.3265 | 0.9994 | 2.7572 | 0.2975 | 0.2978 |
| SAU-FNO | 0.5336 | 0.4086 | 0.9991 | 2.6318 | 0.3779 | 0.4282 |
| DeepOHeat | 3.7055 | 2.9376 | 0.9593 | 10.2608 | 2.8308 | 3.6447 |
| Therm-FM T | 0.5752 | 0.4102 | 0.9986 | 3.1207 | 0.3691 | 0.3992 |
| Therm-FM B | 0.4610 | 0.3198 | 0.9990 | 2.7091 | 0.2870 | 0.3114 |
| Therm-FM L | 0.3536 | 0.2477 | 0.9995 | 2.1448 | 0.2164 | 0.2378 |
Table 15. Complete S2 Within-Family Generalization results under represented layout variation. Errors are in kelvin; higher is better and lower is better for all other metrics.
| Method | RMSE | MAE | MaxAE | Top-50 MAE | ||
| FNO | 1.2912 | 0.9939 | 0.9977 | 5.4362 | 0.6361 | 0.6929 |
| U-Net | 1.2713 | 0.9764 | 0.9978 | 7.0174 | 0.8183 | 0.8383 |
| U-FNO | 0.9391 | 0.7237 | 0.9988 | 5.0106 | 0.5180 | 0.5727 |
| SAU-FNO | 0.9532 | 0.7394 | 0.9988 | 4.1366 | 0.6504 | 0.6826 |
| DeepOHeat | 4.7394 | 3.9033 | 0.9699 | 11.7546 | 4.7950 | 5.9236 |
| Therm-FM T | 1.0496 | 0.7940 | 0.9985 | 4.8432 | 0.5989 | 0.6426 |
| Therm-FM B | 0.8462 | 0.6180 | 0.9990 | 4.2598 | 0.4095 | 0.4614 |
| Therm-FM L | 0.8113 | 0.5852 | 0.9991 | 4.0345 | 0.3696 | 0.4314 |
Table 17. Complete S4 Boundary results under Within-Family Generalization.
Reading S5 zero-shot. Figure 9 shows the strongest frozen model, Therm-FM T, and it is still qualitatively wrong on four of the five compound cross-package OOD cases. The failure is structural rather than numerical: on Case 11 the hot center of the 80-chiplet array is missed almost entirely and the field is predicted near the bottom of the temperature range, on Case 12 only part of the hot region is recovered, on Case 13 the scattered micro-chiplets are merged into a single saturated hot plateau, and on Case 15 the stacked dies collapse into one oversized hot block. Case 14 is the exception: the hot spot of the dominant chiplet is predicted with the correct position and footprint, but the heat spreading around it is overestimated, leaving a ring-shaped error. The error maps reach 10–21 K and follow the misplaced structures instead of the chiplet edges. This is the visual counterpart of the order-of-magnitude jump in Table 18: a predictor that is accurate across diverse layouts, materials, and cooling conditions of the represented families does not carry a usable thermal mapping to an unseen package structure, and no baseline in the benchmark is exempt from this.
Figure 9. S5 zero-shot visualization with Therm-FM T, the lowest-RMSE frozen model, on the held-out Cases 11–15 without target labels (the median-RMSE sample of each case). The predictor misses, merges, or inflates hot regions in four of the five unseen packages, matching the structural failure quantified in Table 19.
| Method | Case11 | Case12 | Case13 | Case14 | Case15 |
| U-FNO | 59.79 | 6.21 | 15.68 | 19.17 | 39.91 |
| SAU-FNO | 55.46 | 4.71 | 37.18 | 9.88 | 23.38 |
| FNO | 94.58 | 20.02 | 50.96 | 20.19 | 25.12 |
| U-Net | 16.17 | 3.25 | 40.36 | 17.44 | 17.05 |
| DeepOHeat | 4.79 | 20.76 | 10.01 | 44.63 | 39.69 |
| Therm-FM T | 8.42 | 9.94 | 11.17 | 7.50 | 6.33 |
| Therm-FM B | 4.48 | 15.33 | 9.03 | 18.07 | 18.36 |
| Therm-FM L | 6.15 | 13.05 | 12.76 | 18.80 | 25.92 |
Table 19. Per-case RMSE (K) on the official S5 zero-shot test. The best value in each held-out case is bold.
Reading S5 after adaptation. Figure 10 applies the same protocol to U-FNO, the strongest operator baseline at , after 100 labeled samples per OOD case. With 500 target labels in total, the global thermal feature is recovered on all five packages: the chiplet array of Case 11, the hot region of Case 12, the dominant die of Case 14, and the stacked dies of Case 15 are placed and scaled correctly, and the scattered micro-chiplets of Case 13 largely reappear as separate hot spots at the correct positions. The error scale drops from 10–21 K to 3.5–5.8 K at the worst pixel and to below one kelvin over most of the field. The remaining residuals are again edge-localized, approaching the error pattern observed in the S2–S4 within-family evaluations; Case 12, which reaches the highest temperatures, retains the largest residual (1.93 K case-mean RMSE, versus 0.74–0.89 K on the other four cases). Target-case adaptation is therefore a practical remedy for cross-package OOD failure—limited target-domain labels can move an otherwise poor zero-shot prediction substantially toward the S2–S4 accuracy regime—but this recovery requires target labels and is reported separately from, not as a substitute for, zero-shot cross-package generalization.
Benchmark Extension Protocol
IC-ThermBench reduces extension effort through reusable data and model interfaces. New datasets compatible with the current tensor representation and split convention can be supplied as input/output MAT pairs and selected by dataset name, reusing the shared loading, normalization, and evaluation routines. A conventional PyTorch predictor requires a model implementation and a MODEL_ZOO entry to use the shared training and evaluation pipeline, rather than a separate model-specific workflow.
Extensions involving new simulators, measurements, or temporal representations may require task-specific conversion, configuration, or adapters, but can build on these shared interfaces. Each extension should publish its case manifest, channel semantics, label provenance, fixed split identifiers, and reference metrics, and specify whether it broadens a Generalization Scope or defines a separate domain-shift track, while preserving the primary benchmark tracks.
Figure 7. S3 visualization with U-FNO. Columns correspond to Cases 1, 3, 5, 7, and 9, with the median-RMSE test sample shown for each case. Rows show ground truth, prediction, and absolute error in kelvin.
Figure 8. S4 visualization with U-FNO. Columns correspond to Cases 1, 3, 5, 7, and 9, with the median-RMSE test sample shown for each case. Rows show ground truth, prediction, and absolute error in kelvin.
References
- [1]Zhuomin Chai, Yuxiang Zhao, Wei Liu, Yibo Lin, Runsheng Wang, and Ru Huang. 2023. CircuitNet: An Open-Source Dataset for Machine Learning in VLSI CAD Applications with Improved Domain-Specific Evaluation Metric and Learning Strategies. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 42, 12 (2023), 3504–5047. doi:10.1109/TCAD.2023.3287970DOI
- [2]Liang Chen, Wentian Jin, and Sheldon X.-D. Tan. 2022. Fast Thermal Analysis for Chiplet Design Based on Graph Convolution Networks. In 27th Asia and South Pacific Design Automation Conference. 485–492.DOI
- [3]Vidya A. Chhabria, Vipul Ahuja, Ashwath Prabhu, Nikhil Patil, Palkesh Jain, and Sachin S. Sapathnerkar. 2021. Thermal and IR Drop Analysis Using Convolutional Encoder-Decoder Networks. In 26th Asia and South Pacific Design Automation Conference. 690–696. doi:10.1145/3354885.3431583arxiv.org/abs/2009.09009
- [4]Animesh Basak Chowdhury, Benjamin Tan, Ramesh Karri, and Siddharth Garg. 2021. OpenABC-D: A Large-Scale Dataset for Machine Learning Guided Integrated Circuit Synthesis. arXiv:2110.11292 [cs.LG]. https://arxiv.org/abs/2110.11292
- [5]Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. ImageNet: A Large-Scale Hierarchical Image Database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE, Miami, FL, USA, 248–255. doi:10.1109/CVPR.2009.5206848DOI
- [6]Yue Hua, Zhi-Qiao Wang, Xin-Yi Yuan, Yu Bai Li, Wei-Tao Wu, and Nadine Aubry. 2023. Estimation of Steady-State Temperature Field in Multichip Modules Using Deep Convolutional Neural Network. Thermal Science and Engineering Progress 40 (2023), 101755. doi:10.1016/j.tsep.2023.101755
- [7]Zhen Huang, Hangyang Wen, Wenkai Yang, Muxi Tang, Depeng Xie, Ting-Jung Lin, Yu Zhang, Wei W. Xing, and Lei He. 2025. Self-Attention to Operator Learning-Based 3D-IC Thermal Simulation. In 62nd ACM/IEEE Design Automation Conference. 263:1–263:7. doi:10.1109/DAC63849.2025.1132988arxiv.org/abs/2510.15968
- [8]Zhen Huang, Haiyang Xin, Dongming Jiang, Yangbo Wei, Hong Wang, Wenkai Yang, Yu Zhang, Wei W. Xing, Ting-Jung Lin, and Lei He. 2026. From Fluid Dynamics to Chip Design: PDE Foundation Model Address Data Bottleneck in 3D-ICs Thermal Simulation. In 63rd ACM/IEEE Design Automation Conference. doi:10.1145/3770743.3803993
- [9]Zhen Huang, Haiyang Xin, Wenkai Yang, Yangbo Wei, Zhiping Yu, Yu Zhang, Wei W. Xing, Ting-Jung Lin, and Lei He. 2026. Therm-FM: Foundation Model Is All You Need for 3D-ICs Thermal Simulation. arXiv:2605.022663
- [10]Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik R. Narasimhan. 2024. SWE-bench: Can Language Models Resolve Real-World GitHub Issues? In International Conference on Learning Representations. OpenReview.net, Vienna, Austria, 24 pages. https://openreview.net/forum?id=VTF8yNQM66
- [11]Wentian Jin, Sherrif Sadiqbatcha, Jinwei Zhang, and Sheldon X.-D. Tan. 2020. Full-Chip Thermal Map Estimation for Commercial Multi-Core CPUs with Generative Adversarial Learning. In IEEE/ACM International Conference on Computer-Aided Design. doi:10.1145/3400302.3415764DOI
- [12]Scott Ladenheim, Yi-Chung Chen, Milan D. Mihajlovic, and Vasilis F. Pavlidis. 2018. The MTA: An Advanced and Versatile Thermal Simulator for Integrated Systems. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 37, 12 (2018), 3123–3136. doi:10.1109/TCAD.2018.2879729
- [13]Yu-Min Lee, Shun-Ping Huang, and Ying-Hui Lin. 2025. Therunet-SP: 2.5D IC Thermal Simulation Using U-Net Convolution and Scaling Procedure. In 26th International Symposium on Quality Electronic Design. IEEE. doi:10.1109/ISQED65160.2025.11014417DOI
- [14]Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. 2021. Fourier Neural Operator for Parametric Partial Differential Equations. In International Conference on Learning Representations.
- [15]Yu-Sheng Lin, Li-Song Lin, Chin-Jui Chang, Ting-Yu Lin, Ya-Wen Yu, Shih-Hong Pan, Wei Cheng Lee, Kai-En Yang, Yi-Chen Lin, Tai-Yu Chen, and Jason Yeh. 2023. Fast SoC Thermal Simulation with Physics-Aware U-Net. In NeurIPS Workshop on Machine Learning and the Physical Sciences. https://neurips.cc/virtual/2023/76089
- [16]Ziyue Liu, Yixing Li, Jing Hu, Xinling Yu, Shinyu Shiau, Xin Ai, Zhiyu Zeng, and Zheng Zhang. 2023. DeepOHeat: Operator Learning-Based Ultra-Fast Thermal Simulation in 3D-IC Design. In 60th ACM/IEEE Design Automation Conference. 1–6. doi:10.1109/DAC56929.2023.10247998DOI
- [17]Lu Lu, Pengzhan Jin, Guofei Pang, Zongqian Zhang, and George Em Karniadakis. 2021. Learning Nonlinear Operators via DeepONet Based on the Universal Approximation Theorem of Operators. Nature Machine Intelligence 3 (2021), 218–229. doi:10.1038/s42256-021-00302-5
- [18]Yao Lu, Zhicheng Guo, Qijun Zhang, Shang Liu, Wenji Fang, Wenkai Li, and Zhiyao Xie. 2024. COOL: A Cooling-aware Point Transformer Framework for Thermal Prediction in Advanced 3D/3.5D IC Packaging. In 63rd ACM/IEEE Design Automation Conference. doi:10.1145/3770743.3804429
- [19]Jingyu Pan, Chen-Chia Chang, Zhiyao Xie, Yiran Chen, and Hai Helen Li. 2024. EDALearn: A Comprehensive RTL-to-Signoff EDA Benchmark for Democratized and Reproducible ML for EDA Research. In Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design. 1–8. doi:10.1145/3667366.3697116arxiv.org/abs/2312.01674
- [20]Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. 2019. Physics-Informed Neural Networks: A Deep Learning Framework for Solving Forward and Inverse Problems Involving Nonlinear Partial Differential Equations. J. Comput. Phys. 378 (2019), 686–707. doi:10.1016/j.jcp.2018.10.045DOI
- [21]Rishikesh Ranade, Haiyang He, Jay Pathak, Norman Chang, Akhilesh Kumar, and Jimin Wen. 2022. A Thermal Machine Learning Solver for Chip Simulation. In Proceedings of the 2022 ACM/IEEE Workshop on Machine Learning for CAD. 111–117. doi:10.1145/3551901.3556484arxiv.org/abs/2209.04741
- [22]Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-Net: Convolutional Networks for Biomedical Image Segmentation. In Medical Image Computing and Computer-Assisted Intervention. 234–241. doi:10.1007/978-3-319-24574-4_28
- [23]Yanliang Sha, Chengxin Zhang, and Quan Chen. 2026. PI-ONet: A Physics-Informed Operator Network for Efficient Thermal Analysis of Multilayer Chiplets. IEEE Transactions on Components, Packaging and Manufacturing Technology 16, 3 (2026), 493–505.DOI
- [24]Michael Joseph Smith, Shengyun Hwang, Vinicius Cabral Do Nascimento, Qiang Qiu, Cheng-Kok Koh, Qanesh Subbarayan, and Dan Jiao. 2023. Real-Time Precision Prediction of 3-D Package Thermal Maps via Image-to-Image Translation. In 2023 IEEE 32nd Conference on Electrical Performance of Electronic Packaging and Systems. 1–3. doi:10.1109/EPEPS58208.2023.10314947
- [25]Makoto Takamoto, Timothy Praditia, Raphael Leiteritz, Daniel MacKinlay, Francesco Alessiani, Dirk Pflüger, and Mathias Niepert. 2022. PDEBench: An Extensive Benchmark for Scientific Machine Learning. In Advances in Neural Information Processing Systems, Vol. 35. Curran Associates, Inc., Red Hook, NY, USA, 14 pages.arxiv.org/abs/2210.07182
- [26]Zhuofuo Tao, Yichen Shi, Yiru Huo, Rui Ye, Zonghang Li, Li Huang, Chen Wu, Na Bai, Zhiping Yu, Ting-Jung Lin, and Lei He. 2024. AMSNet: Netlist Dataset for AMS Circuits. In 2024 IEEE LMl Aided Design Workshop (LAD). 1–5. doi:10.1109/LAD62441.2024.10691781
- [27]Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019. GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding. In International Conference on Learning Representations. OpenReview.net, New Orleans, LA, USA. https://openreview.net/forum?id=rJ4kmR5t7
- [28]Mingyu Wang, Yuanqing Cheng, Weihe Zeng, Zhenjie Lu, Vasilis F. Pavlidis, and Wei W. Xing. 2024. ARO: Autoregressive Operator Learning for Transferable and Multi-Fidelity 3D-IC Thermal Analysis with Active Learning. In IEEE/ACM International Conference on Computer-Aided Design. doi:10.1145/3676536.3676713
- [29]Qipan Wang, Xueqing Li, Tianyu Jia, Yibo Lin, Runsheng Wang, and Ru Huang. 2024. ATPlace2.5D: Analytical Thermal-Aware Chiplet Placement Framework for Large-Scale 2.5D-IC. In IEEE/ACM International Conference on Computer-Aided Design. 39:1–39:9. doi:10.1145/3676536.3676648DOI
- [30]Zhihai Wang, Zijie Geng, Zhaojie Tu, Jie Wang, Yuxi Qian, Zhexuan Xu, Ziyan Liu, Siyuan Xu, Zhentao Tang, Shixiong Kai, Mingxuan Yuan, Jianye Hao, Bin Li, and Feng Wu. 2025. Benchmarking End-to-End Performance of AI-Based Chip Placement Algorithms. In Advances in Neural Information Processing Systems 38. 22670–22696. doi:10.52202/085713-0673
- [31]Gege Wen, Zongyi Li, Kamyar Azizzadenesheli, Anima Anandkumar, and Sally M. Benson. 2022. U-FNO: An Enhanced Fourier Neural Operator-Based Deep-Learning Model for Multiphase Flow. Advances in Water Resources 163 (2022), 104180. doi:10.1016/j.advwatres.2022.104180
- [32]Jimin Wen, Stephen Pan, Norman Chang, Wen-Tze Chuang, Wenbo Xia, Deqi Zhu, Akhilesh Kumar, En-Chi Yang, Karthik Srinivasan, and Ying-Shun Li. 2020. DNN-Based Fast Static On-Chip Thermal Solver. In 2020 36th Semiconductor Thermal Measurement, Modeling & Management Symposium (SEMI-THERM). 65–75. doi:10.23919/SEMI-THERM50369.2020.9142855
- [33]Chun Yuan Yang, Shijun Wang, Xiang Gao, Zhong Zhang, and Zhiping Xu. 2026. Real-Time 3D Thermal Prediction for Advanced Packaging via Machine Learning. International Journal of Heat and Mass Transfer 267 (2026), 128936. doi:10.1016/j.ijheatmasstransfer.2026.128936DOI
- [34]Ziyao Yang, Jingbo Sun, Vidya A. Chhabria, and Yu Cao. 2025. Adaptive Graph Learning for Efficient Thermal Analysis of Multi-Stacking Chiplet Systems under Interface Variations. In IEEE/ACM International Conference on Computer-Aided Design. 1–9. doi:10.1109/ICCAD62669.2025.11240700DOI
- [35]Xinling Yu, Ziyue Liu, Hai Li, Yixing Li, Xin Ai, Zhiyu Zeng, Ian Young, and Zheng Zhang. 2026. DeepOHeat-v1: Efficient Operator Learning for Fast and Trustworthy Thermal Simulation and Optimization in 3D-IC Design. IEEE Transactions on Components, Packaging and Manufacturing Technology 16, 1 (2026), 82–96.
- [36]Bingrui Zhang, Wei W. Xing, Xin Zhao, and Yiquan Sun. 2025. T-Fusion: Thermal Modeling of 3D ICs with Multi-Fidelity Fusion. In 30th Asia and South Pacific Design Automation Conference. 1406–1412. doi:10.1145/3658617.3679749
- [37]Dekang Zhang, Dan Niu, Zhiyu Cao, Yichao Dong, Jingweijia Tan, and Changyin Sun. 2025. Therm-PCT: A Geometry-Material Aware Point Cloud Transformer for Large-Scale Unstructured Thermal Analysis in 2.5D ICs. In IEEE/ACM International Conference on Computer-Aided Design. 1–9. doi:10.1109/ICCAD62669.2025.11241008
- [38]Tianxiang Zhu, Qipan Wang, Yibo Lin, Runsheng Wang, and Ru Huang. 2024. FaSt-Therm: Fast and Stable Full-Chip Transient Thermal Predictor Considering Nonlinear Effects. In IEEE/ACM International Conference on Computer-Aided Design. 144:1–144:9. doi:10.1145/3676536.3676738
- [39]Tianxiang Zhu, Qipan Wang, Yibo Lin, Runsheng Wang, and Ru Huang. 2024. FaSt-Therm: Fast and Stable Full-Chip Transient Thermal Predictor Considering Nonlinear Effects. In IEEE/ACM International Conference on Computer-Aided Design. 144:1–144:9. doi:10.1145/3676536.3676738DOI






