2D-ThermAl: Physics-Informed Framework for Thermal Analysis of Circuits using Generative AI

TL;DR

This work tackles the challenge of rapid, accurate circuit-scale thermal analysis suitable for early-stage IC design. It introduces ThermAl, a physics-informed conditional generative U-Net that predicts full-chip transient and steady-state thermal maps from activity profiles, guided by a CV-based heat conduction regularizer. The approach delivers RMSE ≈ 0.71°C on 256×256 grids with millisecond inference, achieving up to ~200× speedups over traditional FEM while maintaining physical plausibility. Extended-range cross-validation shows robust generalization to higher temperatures (25–95°C), and ablation studies confirm the benefits of physics regularization, source-target conditioning, and feature-level concatenation. ThermAl thus provides a practical, scalable tool for hotspot detection and thermal-aware design exploration in modern EDA workflows.

Abstract

Thermal analysis is increasingly critical in modern integrated circuits, where non-uniform power dissipation and high transistor densities can cause rapid temperature spikes and reliability concerns. Traditional methods, such as FEM-based simulations offer high accuracy but computationally prohibitive for early-stage design, often requiring multiple iterative redesign cycles to resolve late-stage thermal failures. To address these challenges, we propose 'ThermAl', a physics-informed generative AI framework which effectively identifies heat sources and estimates full-chip transient and steady-state thermal distributions directly from input activity profiles. ThermAl employs a hybrid U-Net architecture enhanced with positional encoding and a Boltzmann regularizer to maintain physical fidelity. Our model is trained on an extensive dataset of heat dissipation maps, ranging from simple logic gates (e.g., inverters, NAND, XOR) to complex designs, generated via COMSOL. Experimental results demonstrate that ThermAl delivers precise temperature mappings for large circuits, with a root mean squared error (RMSE) of only 0.71°C, and outperforms conventional FEM tools by running up to ~200 times faster. We analyze performance across diverse layouts and workloads, and discuss its applicability to large-scale EDA workflows. While thermal reliability assessments often extend beyond 85°C for post-layout signoff, our focus here is on early-stage hotspot detection and thermal pattern learning. To ensure generalization beyond the nominal operating range 25-55°C, we additionally performed cross-validation on an extended dataset spanning 25-95°C maintaining a high accuracy (<2.2% full-scale RMSE) even under elevated temperature conditions representative of peak power and stress scenarios.

Figures (8)

• Figure 1: As transistor density in modern microprocessors increases, traditional thermal estimation methods often struggle with complex heat dissipation, leading to inaccuracies in temperature profiling. While commercial FEM tools offer high fidelity, they can be too slow and computationally inefficient for full-chip modeling. Simultaneously, thermal cameras tend to lack the high-resolution detail needed for analysis at the circuit level. ThermAl tackles these limitations by using a generative neural network to deliver fast and accurate thermal estimation. More importantly, it integrates easily into the design exploration stage of the GDS generation, allowing designers to identify and correct thermal hotspots early on. Thermal-sensitive flow can prevent numerous iterations of designs, induced by thermal failures discovered late in the process, and thus simplify the overall circuit optimization process.
• Figure 2: Workflow for Dataset Generation: Circuit designs and power distribution data are first generated using standard EDA tools. These, along with specified thermal parameters, are input into FEM-based time-stepped simulations that capture transient heat flow, producing thermal maps at various intervals. The resulting comprehensive dataset serves as ground truth for training ThermAl, enabling accurate predictions across diverse circuit configurations and thermal conditions.
• Figure 3: (a) Comprehensive library of steady-state and transient thermal behaviors across diverse workloads and initial conditions, generated using the commercial finite element-based Multiphysics software (COMSOL). (b) Transient snapshots illustrate the time-evolving evolution of heat flow, highlighting changes in temperature profiles over time within the circuit.
• Figure 4: Overview of the ThermAl framework: A novel hybrid U-Net network combining CNN and positional embedding for feature extraction and processing. The network accepts a pair of source-target images $\{E: \text{source}, E^\prime: \text{target}\}$, along with a input query image, $\{I\}$, and processes them through separate UNet-based encoders to extract multi-scale features. The resulting feature maps are concatenated to form $f_c$, which is then augmented with positional embeddings $f_{pos}$ and fed into a decoder to generate a dense prediction. The framework is trained using a composite loss function that combines the RMSE loss with a physics-aware regularizer, ensuring that predictions adhere to the underlying heat conduction dynamics. The flow-diagram illustrates how the model predicts transient heat dissipation and thermal flow over various time intervals, capturing the temporal evolution of thermal behavior.
• Figure 5: Qualitative results of ThermAl: The image demonstrates the generation of thermal maps from an initial sample (leftmost column), showing the progression of heat distribution over time (intermediate output columns). These steps capture the dynamic heat flow over time steps $t = 1, 5, 10, 20, 50$, with the final predicted output map at $t = 100$ compared against the ground truth (GT) in the rightmost columns. The error maps highlight the pixel-wise differences between the predicted and true simulated thermal maps.