AI-Driven Simulation for Faster and Confident Thermal Design

This blog is based on the presentation delivered by Jiwon Yu from Celestica titled ‘ Validating and Accelerating Thermal Design Optimization using AI-Driven Simulation’ at CadenceLIVE Silicon Valley 2026

As thermal architecture grows in complexity, conventional trial-and-error approaches become too slow to keep pace with design demands. Engineers need faster ways to evaluate trade-offs, optimize performance, and converge on the right solution early in the design cycle. This blog presents a Celestica-led study that used Cadence Optimality Intelligent System Explorer and Celsius EC Solver to validate and accelerate AI-driven thermal simulation workflows for complex electronics cooling.

The results showed that AI-guided optimization can compress workflows that once required months into just days, dramatically improving engineering productivity. A key result was that the AI-driven approach reached a comparable optimum even when starting from a suboptimal initial design, providing confidence that it can guide thermal design decisions earlier in the development process.

Why Thermal Teams Need a Different Optimization Workflow

The motivation is straightforward: fixing design issues early is far less costly than addressing them after production. Yet high-performance systems keep getting more complex, with tightly coupled interactions among airflow, component placement, and thermal behavior.

That complexity is pushing conventional trial-and-error design to its limits. Engineers spend significant time building detailed models and running repeated simulations to balance competing requirements. In this study, manual optimization stretched across several months while program timelines remained fixed. The need is clear: a faster, more effective approach.

What This Benchmark Study Was Designed to Prove

This benchmarking study by the Celestica team evaluated AI as a reliable engineering partner by comparing manual trial-and-error optimization with Cadence Optimality Intelligent System Explorer, an AI-driven multiphysics optimization software that drives productivity by exploring the full design space, on the same product.

The System: A High-Power Network Switch with Tight Thermal Trade-Offs

The use case was a 2U network switch with 64 Octal Small Form-factor Pluggable modules (OSFP), a Tomahawk 5 device, and a Computer-on-Module Express (COMe) module positioned behind the OSFP modules. The 100% air-cooled platform carried about 3 kW of total thermal load and was designed for a 45°C ambient environment.

The simulation model was built in Celsius EC solver and ran at 70 million grid cells. Air is designed to flow from front to back, and the Tomahawk 5 and COMe module receive air preheated by the OSFP modules, creating the central thermal tradeoff in the design.

The Design Problem Was Not a Single Hotspot but a Balance Problem

The core challenge centered on the frontal open area. Reducing it increases airflow velocity through the OSFP modules and helps cool them, but it also reduces fresh air to downstream components, raising temperatures on the Tomahawk 5 and CPU-related hardware.

The goal was to find the best thermal balance across these conflicting requirements. Reaching that balance manually took several months, making the design a strong candidate for automated optimization.

How the AI-Driven Optimization Study Was Set Up

To balance these thermal objectives, the study defined five input variables: front opening area, Tomahawk 5 heat-sink fin thickness, number of fins, and COMe module position along the x- and z-axes, all within mechanical limits. Because the temperature targets competed with one another, the team used a global objective function rather than a single metric. OSFP temperatures were weighed five times more heavily because they were harder to keep within specification, and each temperature was constrained to remain below its specified limit.

The optimization setup followed design-space navigation. Iteration count was set to two times the number of input parameters, and design of experiment (DOE) iteration count to three times the number of input parameters. In this case, each iteration used 24 CPU cores, with no parallel DOE jobs.

Testing Whether AI Depends on the Starting Design

To test how much the starting point affected the outcome, the study used two initial conditions: a semi-optimized design created by a thermal engineer through trial and error, and a not-optimized design with arbitrarily selected variables. This made it possible to assess whether Optimality could improve performance from different starting conditions and still reach the same optimal configuration.

What the Optimization Data Showed

The raw optimization data captured the full range of studied variables and outcomes. Within that dataset, Optimality identified one scenario as the most efficient design and minimized the global objective function across a broad design space.

Computational fluid dynamics (CFD) convergence in Celsius EC Solver was also monitored, and the temperature values converged well, supporting the reliability of the results.

The analysis focused on convergence, correlation, and sensitivity. The convergence plots showed how Optimality navigated the design space and systematically improved thermal performance.

Sensitivity analysis identified front opening area, heat-sink fin thickness, and number of fins as the three most important variables.

Most importantly, comparing Optimality predictions with full 3D CFD results yielded a correlation of 0.93, demonstrating that the AI-driven predictions were both accurate and reliable. Rather than relying on thousands of computationally heavy full 3D CFD simulations, Optimality uses AI-guided design space exploration and surrogate modeling to learn the relationship between design parameters and thermal performance.

This allows it to evaluate and rank potential design alternatives using far fewer CFD runs, dramatically reducing optimization time while still converging on solutions that closely match full-physics results.

What Changed When AI Was Compared Directly with Manual Optimization

Manual optimization was already strong, which makes the comparison meaningful. It scored 511.9 on the objective function, while the AI-driven approach reached 509.36, outperforming the manual solution on a difficult thermal balancing problem.

A key result was consistency across starting inputs. From the semi-optimized design, Optimality reached 509.36. From the not-optimized design, it reached 509.622. That near match suggests it can arrive at essentially the same optimal configuration from more crudely selected starting conditions. The optimization also indicated that, at this airflow, AI favored a thicker heat sink with fewer fins for the Tomahawk 5 than manual optimization did.

Why This Matters for Future Engineering Workflows

Taken together, the results show why this approach matters for future engineering workflows. It is accurate and reliable, with a correlation greater than 0.9 against full 3D CFD simulations; efficient, cutting months of repeated trial and error; and practical for shortening design time in next-generation server and networking systems. Its ability to reach nearly the same optimum from different starting points also supports using it earlier in the design process.

Just as importantly, it lets engineers spend less time on repeated setup-and-check loops and more time making more fundamental design decisions.

To learn more about validating and accelerating thermal design optimization with AI-driven simulation, watch the on-demand webinar on the CadenceLIVE Silicon Valley 2026 events page.