Physics-Based Validation: The Foundation Of Trustworthy Agentic EDA

This is the third article in a series examining what it takes to build production-ready AI agents for EDA. The first explored the architectural decisions required for agents to operate reliably in production environments. The second examined what self-verifying means in an agentic EDA workflow: how agents continuously check their decisions as they work, rather than waiting until the end of a run.

But what exactly are those decisions being verified against?

The answer is physics. More specifically, the deterministic, physics-based EDA engines that define whether a design is correct. A self-verifying agent checks its own work, and a physics-validated workflow confirms that work is correct.

Semiconductor design is governed by physics

Chip design has always faced physics-based constraints. Every transistor, every wire, every timing path, and every thermal gradient is governed by the fixed laws of physics, regardless of how the design was created. This is why tapeout is binary: silicon either works or it does not.

When AI enters this domain, that physical reality does not change. The relevant question is not whether an AI system can reason about physics, but whether the outputs it produces have been confirmed by the engines that enforce that physics.

That distinction separates a model that is capable of reasoning about physics from a workflow that is physics-validated.

What physics-validated means

A physics-validated workflow is one in which every agent decision is confirmed against a deterministic, physics-based EDA engine before the next step begins. Validation does not happen only at the end of a workflow or after an entire run is complete. It happens continuously, at every step, against the engines engineers have trusted for decades.

The agent executes a step, and the engine determines whether the result meets the design’s physical constraints. If it does not, the agent identifies the issue, resolves it, and continues. Physics-validated means the engine, not the model, is the authority on correctness.

Why it matters at tapeout

In semiconductor design, there is zero margin for unverified decisions. Tapeout dates are fixed, respins are expensive, and missed market windows do not come back.

For engineering organizations deploying agentic AI at scale, the key question is not simply whether the workflow is fast. It is whether the outcomes it produces can be trusted without requiring engineers to manually re-verify every autonomous decision.

An agentic workflow that reasons correctly about physics but does not validate its outputs against physics-based engines introduces risk at every autonomous step. The longer a workflow runs without validation, the further a design can drift from correctness before an error surfaces.

In a long-running autonomous workflow that executes across multiple tools, physical domains, and hours without human intervention, unvalidated decisions can compound. Continuous physics-based validation prevents those errors from propagating through the workflow.

3D-IC design: A multi-physics challenge

In Part 2, I described an autonomous library characterization workflow in which the agent continuously checks its work against physics-based EDA engines. That workflow demonstrates the principle at a specific task level. The need for continuous validation increases when a workflow spans multiple tools and physical domains, and 3D-IC design shows why. It is the ultimate test of physics, and the clearest example of why semiconductor design is physics.

AI factories require more compute, memory bandwidth, and thermal performance within an increasingly constrained physical footprint. As power, performance, and area scaling on monolithic dies reaches practical limits, chiplet-based 2.5D and 3D architectures allow compute and memory to scale through tightly integrated packages.

That level of integration makes the physical design problem considerably more complex. Stacked active dies make heat removal and temperature interactions harder to predict, while packaging can introduce mechanical stress. Electrical, power, thermal, and mechanical effects are interconnected, so a change in one area can affect the others. Addressing these effects through manual, fragmented, and tool-intensive workflows creates frequent engineering handoffs, slows iteration, and limits the number of design alternatives teams can evaluate.

An autonomous workflow for 3D-IC design

Siemens has developed an agentic engineering workflow that autonomously coordinates 3D-IC design and analysis. MCP servers and skill files for different EDA tools are connected to Fuse EDA AI Agent, which acts as an MCP aggregator that can be connected to any coding agent of choice. In this particular example, Claude Code is used as the agentic coding IDE, orchestrating a nine-stage electro-thermal workflow spanning seven Siemens EDA tools, running 100 to 200 agentic turns over more than 12 hours.

Each stage performs a distinct engineering task, from floor planning, library characterization, and power analysis to thermal signoff, system-level thermal simulation, parasitic extraction, and power- and temperature-aware SPICE simulation. Every stage is validated using signoff-grade, physics-based solvers.

The workflow runs autonomously, executing analyses, evaluating results, identifying missed targets, and continuing to refine the design. Teams can sweep heat transfer coefficient (HTC) boundaries, power budgets, and material stack-ups across 200 to 500+ design points per cycle. This enables teams to configure tools up to 20 times faster, reduce engineering effort by up to 95%, and evaluate up to 50 times more design options than a conventional manual flow.

The demo below shows the electro-thermal workflow in action:

How do I know if my workflow is physics-validated?

When evaluating an agentic EDA workflow, one question determines whether its architecture is physics-validated: At each step, is the agent’s output confirmed by a deterministic, physics-based EDA engine before the workflow proceeds?

If the answer is yes, the workflow is physics-validated. The engine is the arbiter, and every autonomous decision is confirmed against the physical reality of the design.

Defining trusted autonomy in EDA

Taken together, the three articles in this series define the requirements for production-ready agentic EDA: an architecture designed for real engineering environments, a self-verifying mechanism that continuously checks agent decisions, and a physics-validated standard that determines whether those decisions are correct.

This is the architectural standard Siemens EDA has built into the Fuse EDA AI Agent. The result is trusted autonomy: validated engineering outcomes at every step of the workflow.

Read the first two articles in the series: