🔬 materials science

AI for Science with GPT-6 Astra: Thermal Design and Electrothermal Analysis of 2D CFET

This paper demonstrates that an AI agent workflow utilizing GPT-6 Astra can successfully optimize the thermal design of 2D CFET inverters by proposing and validating structural modifications that significantly reduce peak temperatures while quantifying associated electrical trade-offs.

Original authors: Min-Hui Kim, Khushi Sharma, Sarah Zhang, Ye Wang

Published 2026-09-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Min-Hui Kim, Khushi Sharma, Sarah Zhang, Ye Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Modern electronics are shrinking to a point where the very act of switching a computer on and off generates a surprising amount of heat. When engineers stack different types of transistors on top of one another to save space, they trap that heat in a tiny footprint, much like piling blankets over a warm body. The materials between these layers act as insulators, making it difficult for the warmth to escape. If this heat builds up, it changes how the electricity flows, which in turn creates more heat, potentially causing the device to fail. For the next generation of ultra-small chips, finding a way to pull this heat away without making the device larger or slower is one of the most critical challenges in physics and engineering.

To tackle this, a team of researchers recently turned to a new kind of helper: an artificial intelligence scientist. They did not simply ask the computer to solve a math problem; they gave it a virtual laboratory and a specific goal. The researchers were studying a type of future chip made from two-dimensional materials, which are sheets of atoms so thin they are essentially flat. These chips are designed to stack n-type and p-type transistors vertically, creating a compact inverter—a basic building block that flips electrical signals. The team wanted to know if an AI could figure out how to rearrange the tiny metal wires inside this chip to cool it down, and whether doing so would hurt the chip's performance.

The AI system, named Astra, was placed inside a digital simulation of the chip. Its task was to test different shapes and positions for the metal wires that connect the top and bottom layers of the device. The researchers set strict rules: the total amount of metal used had to stay exactly the same, and the chip had to handle a specific amount of electrical power. Astra began by testing a hypothesis that seemed logical at the time: perhaps making the wire on the hotter top layer wider would help. However, the simulation showed the opposite. When the AI widened that top wire, the chip actually got hotter. The AI quickly learned that the solution lay not in making things bigger, but in moving them. It proposed a design where the metal on the top layer was narrowed and shifted inward, while the saved metal was moved to the lower layer. This redistribution allowed heat to flow more easily downward.

But the AI did not stop there. A second, coordinating AI agent looked at the results and suggested a new addition: a narrow vertical path, like a chimney, extending from the device down toward the base to carry heat away. When the researchers combined the repositioned wires with this new heat path, the results were clear. In the simulation, the peak temperature of the chip dropped by 1.67 degrees Kelvin compared to the original design. This was a significant improvement, proving that the AI could not only find a better shape but also invent a new way to move heat that a human might not have immediately considered.

The researchers then tested this new design under more realistic conditions, where the chip was actually running a circuit and generating its own heat based on how hard it was working. In this scenario, the cooling effect was smaller but still present, lowering the temperature by about 0.3 degrees Kelvin. More importantly, the study revealed the cost of this cooling. To achieve the lower temperature, the chip's ability to conduct electricity dropped slightly, losing about two percent of its maximum current. This trade-off is crucial for engineers to understand: better cooling is possible, but it comes with a small penalty in speed or power. The AI successfully quantified this balance, showing exactly how much performance was lost to gain a specific amount of cooling.

To ensure these results were not just a glitch in the computer code, the team asked several different AI models to repeat the calculations independently. Most of them agreed perfectly, confirming the findings. However, one model failed dramatically, producing a result that was off by more than 100 degrees, which helped the researchers identify a specific error in how that model handled the data. This process of checking and re-checking gave the team confidence that the 1.67-degree improvement was real within the simulation. The study concludes that while the design has not yet been built in a physical lab, the workflow used here—where an AI proposes a structure, tests it, and measures the cost—works effectively. It demonstrates that artificial intelligence can now act as a partner in scientific discovery, suggesting physical designs and rigorously testing them against the laws of physics before a single piece of metal is ever manufactured.

Technical Summary: AI for Science with GPT-6 Astra: Thermal Design and Electrothermal Analysis of 2D CFET

Problem Statement
Two-dimensional (2D) complementary field-effect transistor (CFET) integration offers strong electrostatic control and dense stacking but concentrates heat within a single footprint. In 2D CFET inverters, dielectrics and thermal boundary resistances impede heat removal, while temperature variations alter transistor currents and redistribute power. Consequently, thermal designs must be evaluated under actual circuit operation rather than solely under prescribed heat loads. The challenge lies in determining which geometries to test, how to compare candidates fairly under fixed constraints (such as metal volume), and how to systematically investigate design trade-offs using computational tools.

Methodology
The authors employ an AI scientist workflow centered on "GPT-6 Astra" to navigate the design space of 2D CFET inverters. The methodology integrates a supplied Python-based electrothermal model with autonomous agents:

Device Model: The simulation uses four MoS₂ sheets (n-channel) stacked above four WSe₂ sheets (p-channel) with a gate length ( $L_g$ ) of 12 nm. The model solves anisotropic steady-state heat conduction via finite volumes, accounting for thermal boundary resistance, fixed bottom temperatures (300 K), and top heat exchange.
Electrical Loading: Two loading conditions are applied: (1) fixed-power tests (20 µW) to isolate thermal resistance effects, and (2) circuit-level operation where power is derived from inverter currents at $V_{DD} = 0.7$ V. The electrical model includes charge-sheet current, temperature-dependent mobility ( $T^{-1.35}$ ), and saturation velocity ( $T^{-0.5}$ ).
Agent Workflow:
- Astra: An agent tasked with searching interconnect geometries. It hypothesizes that widening the hotter upper source interconnect and moving it inward might improve cooling.
- Coordinating Agent: Interprets simulation trends and proposes structural modifications, such as adding substrate-directed vias, when initial hypotheses yield unexpected results.
- Reproduction: Separate agent runs verify code consistency and implementation correctness.

Key Contributions and Results

AI-Driven Structural Optimization: Contrary to the initial hypothesis that widening the upper interconnect would help, Astra identified that narrowing and repositioning the upper source interconnect while reallocating the saved metal to the lower tier was more effective. This redistribution reduced the peak temperature rise by 0.88 K under fixed-power conditions.
Synergistic Design: The coordinating agent proposed adding a narrow via directed toward the substrate to enhance downward heat flow. When combined with the redistributed interconnect, the design achieved a total peak temperature rise reduction of 1.67 K (from a baseline of 141.63 K to 139.96 K) at a fixed metal volume and 20 µW power dissipation. This combined reduction exceeded the sum of the individual changes.
Electrothermal Trade-offs: Under circuit-level operation ( $V_{DD} = 0.7$ V), the combined design lowered the maximum inverter temperature by 0.335 K (from 330.75 K to 330.41 K). When metal resistance was included, the cooling effect was 0.6 K, but this came with a 2.2% loss in nFET on-current, quantifying the electrical cost of improved thermal management.
Sensitivity Analysis: The study revealed that lower temperatures do not always imply better heat removal. Scaling the effective contact length showed that shortening contacts from 20 nm to 8 nm increased thermal resistance by 4.2% yet lowered peak temperature by 6.2 K, primarily because the nFET on-current dropped by 17%, reducing the total power dissipation.
Reproduction and Validation: The workflow included a rigorous reproduction phase. Most AI agents (including GPT-6 Astra and various Claude models) reproduced the thermal reference cases with errors below $10^{-10}$ K. However, one agent (Claude Haiku 4.5) failed significantly (104.95 K error) due to indexing and unit errors, which the authors retained to demonstrate the necessity of verification steps.

Significance and Claims
The paper claims that an AI scientist workflow can successfully propose thermal structures, test them under common physical and circuit constraints, and quantify their electrical costs. The results demonstrate that agents can navigate complex design spaces to find non-intuitive solutions (such as narrowing a hot interconnect) that outperform baseline designs. The authors emphasize that while the workflow produces testable structural hypotheses and evaluates circuit-level trade-offs, experimental validation remains necessary. The study does not claim to have solved the thermal management problem definitively but rather illustrates the capability of AI agents to assist in the hypothesis generation and evaluation phases of semiconductor research. The work remains a computational study, noting that neither the search nor the results have been independently repeated or experimentally validated.

More like this