AI-Era Simulation EDA Evolution: Physics Solvers, EDA Agents, and a Trusted Engineering Loop

Abstract: AI will not replace physics solvers such as SPICE, electromagnetic, and thermal solvers, but it will change how these solvers are invoked and verified. Based on the keynote presentation at ICDIA 2026, this article analyzes the evolution of simulation EDA toward Agentic EDA, as well as the tool interfaces, physical verification, result reproducibility, and permission auditing capabilities required to implement EDA Agents.

From August 20–21, 2026, the China Integrated Circuit Innovation Application Conference & Exhibition (ICDIA 2026) was held at the Nanjing Yangtze River International Conference Center. With the theme “Chip and Machine Advancing Together, Creating a New Era of Intelligent Innovation,” this year’s conference focused on breakthroughs in key integrated circuit technologies, industry-chain collaboration, and the implementation of application scenarios.

Junyong Deng, Deputy General Manager of Julin Technology, was invited to attend the conference’s keynote forum and deliver a speech titled “Where Are Simulation EDA Tools Headed in the AI Era?” He shared his views on the division of responsibilities between AI and physics solvers, the conditions required for EDA Agent implementation, and the future competitiveness of simulation tools. He also introduced Julin Technology’s product roadmap for trusted engineering agents.

AI Is Not Simply About Making a Single Simulation Run Faster

In his presentation, Junyong Deng outlined the changing role of AI in EDA simulation: from the early stage of executing tasks in batches according to s and rules, to using data to assist parameter optimization, and then to embedding machine learning into tools for prediction and parameter tuning. Today, EDA is moving further toward Agentic EDA, in which agents can understand objectives, decompose tasks, call tools, and continue iterating based on the results.

This means that the division of responsibilities between engineers and tools will also change. In the past, engineers needed to configure tasks one by one, wait for results, identify problems, and then modify parameters again. In scenarios with clearly defined boundaries, agents can take over a large amount of repetitive work, while engineers focus more on defining objectives, constraints, acceptance criteria, and points requiring human review.

Based on this, Junyong Deng proposed that when evaluating AI-powered simulation tools, the focus should not only be on how much time a single calculation saves, but also on whether the tool reduces the total cycle time consisting of configuration, waiting, rework, and manual verification.

The truly meaningful metric is whether an engineering team can obtain a validated result that can be used for the next decision more quickly.

AI Expands Exploration, While Physics Solvers Preserve Engineering Truth

Regarding the question of whether “AI will replace traditional simulation tools,” Junyong Deng gave a clear answer: AI will not replace physics solvers, but it will change how solvers are invoked and verified.

In complex design spaces, AI is well suited to rapidly searching parameter combinations, screening candidate solutions, identifying anomalies, and organizing results. It can help engineers explore more possibilities and reduce a large number of low-value repetitive attempts.

However, final signoff results still require deterministic calculations from physics solvers such as SPICE, electromagnetic, and thermal solvers. In particular, when dealing with new structures, new materials, and extreme operating conditions that are not covered by training data, engineering teams must know what inputs produced each result, which tool version and solver settings were used, and whether the same conclusion can be reproduced by running the process again.

Therefore, a more appropriate relationship between the two is an “exploration–verification–iteration” loop: AI proposes candidate solutions and anomaly hypotheses, while physics solvers perform verification. The verification results are then returned to AI for the next round of task adjustment.

AI improves exploration efficiency, while physics solvers provide the basis for final engineering decisions.

EDA Agents Cannot Prove Their Trustworthiness Through “Self-Evaluation”

Junyong Deng emphasized that the trustworthiness of an EDA Agent cannot be established by whether its language sounds like that of an expert. Instead, it depends on whether every action it takes can be verified by deterministic results.

A trusted operating mechanism should first clearly define the engineering objectives, constraints, and acceptance criteria. The Agent then decomposes tasks, selects tools, and performs calculations while recording input data, tool versions, and complete configurations.

After the calculation is completed, a rule engine or physics solver should determine whether the result passes, rather than allowing the large language model to evaluate its own output.

If a task fails, the system should also be able to adjust parameters and rerun the task, roll back to the previous reliable state, or hand the anomaly over to an engineer.

Ultimately, what enters the knowledge base should be verified processes and results, rather than unverified model-generated inferences.

The value of a multi-Agent architecture also comes from this division of responsibilities: an Agent responsible for coordinating objectives, specialized Agents responsible for models and rules, and Agents responsible for executing simulations and generating reports can share the same project state.

Critical stages are then jointly controlled through deterministic verification and human review.

The number of Agents itself does not create value. Clear responsibilities and verifiable results are the prerequisites for an engineering closed loop.

Agent Implementation Starts with the Toolchain

In his presentation, Junyong Deng summarized five fundamental capabilities required for EDA Agent implementation:

1. Tools need to understand EDA objects, constraints, and units;

2. Core functions need to be callable through stable interfaces;

3. Output results need to be verifiable through Abstract: AI will not replace physics solvers such as SPICE, electromagnetic, and thermal solvers, but it will change how these solvers are invoked and verified. Based on the keynote presentation at ICDIA 2026, this article analyzes the evolution of simulation EDA toward Agentic EDA, as well as the tool interfaces, physical verification, result reproducibility, and permission auditing capabilities required to implement EDA Agents.

   From August 20–21, 2026, the China Integrated Circuit Innovation Application Conference & Exhibition (ICDIA 2026) was held at the Nanjing Yangtze River International Conference Center. With the theme “Chip and Machine Advancing Together, Creating a New Era of Intelligent Innovation,” this year’s conference focused on breakthroughs in key integrated circuit technologies, industry-chain collaboration, and the implementation of application scenarios.

   Junyong Deng, Deputy General Manager of Julin Technology, was invited to attend the conference’s keynote forum and deliver a speech titled “Where Are Simulation EDA Tools Headed in the AI Era?” He shared his views on the division of responsibilities between AI and physics solvers, the conditions required for EDA Agent implementation, and the future competitiveness of simulation tools. He also introduced Julin Technology’s product roadmap for trusted engineering agents.

4. rules or physics engines;

5. Tasks need to support pause, retry, rollback, and human takeover;

6. The system also needs local deployment, permission control, audit trails, and IP isolation capabilities.

This also explains why different scenarios will progress at different speeds.

Tasks such as rule checking, model quality checking, impedance exploration, batch simulation, and report generation have clearly defined inputs and outputs and objective acceptance criteria, making them more suitable for establishing closed loops first.

In contrast, root-cause diagnosis for complex electromagnetic problems, non-standard topologies, and cross-physical-domain trade-offs require clearer verification boundaries and greater human involvement.

Therefore, the appropriate evolution path for Agents should start with single-task automation, gradually expanding to single-product closed loops, multi-tool orchestration, and cross-domain collaboration.

Each time the scope of autonomy expands, the system needs to reassess the error rate, result reproducibility, failure rollback, operating cost, and responsibility boundaries, rather than using model accuracy alone to determine whether the system is mature.

Large Companies Compete on Full-Stack Capabilities; Challengers Need to Compete on “Closed-Loop Density”

Junyong Deng pointed out in his presentation that simulation EDA vendors face four practical challenges in the AI era:

• Customer design data cannot be freely aggregated, while failure samples are particularly scarce;

• Model hallucinations and inputs outside the training scope may lead to incorrect judgments;

• Interfaces and data semantics of legacy tools are not unified;

• Once AI enters production workflows, IP isolation, permissions, and responsibility for results must have clear mechanisms.

Based on these constraints, he believes that the future competitiveness of EDA vendors will increasingly come from callable tool capabilities, traceable data pipelines, scenario-based evaluation systems, open tool orchestration methods, and productized mechanisms for permissions, auditing, rollback, and reproducibility.

These capabilities cannot be obtained simply by replacing one large model with another. They require long-term accumulation in real engineering workflows.

For challengers, rather than pursuing a universal Agent that covers every workflow, it is better to select scenarios that are high-frequency, have clear pain points, and deliver measurable benefits.

The goal is to truly connect tool invocation, physical verification, anomaly handling, and human review into a closed loop.

The presentation described this differentiated capability as “closed-loop density.”

Julin Technology’s Approach: Make Products Reliable First, Then Build Trusted Agents

Based on the above assessment, Julin Technology has chosen a product path of “toolchain first, gradual Agent adoption.”

In 2026, Julin Technology will focus on building callable product capabilities and a data evaluation foundation. The company will unify core function interfaces, task deion methods, and result-recording methods, providing existing tools with the engineering foundation required for reliable AI invocation.

In 2027, Julin plans to launch Copilot and Check Agent capabilities for configuration generation, report summarization, and rule checking. The initial focus will be on validating single-scenario closed loops in tasks with clearly defined boundaries.

In 2028, Julin plans to further advance the SI Signoff Agent toward cross-tool collaboration and the generation of candidate Engineering Change Order (ECO) solutions, while improving rollback, reproducibility, and human review mechanisms.

The SI Signoff Agent is the first minimum viable closed loop planned by Julin Technology. It will focus on the signal integrity signoff process and connect design import, link identification, task configuration, solver execution, rule checking, problem localization, ECO candidate generation, and report generation.

The goal is to allow every calculation to have a basis and every conclusion to be reproducible within clearly defined human-review boundaries, rather than simply adding a chat window to an EDA tool.

From “Providing Answers” to “Delivering Validated Results”

At the end of the presentation, Junyong Deng responded to the question “Where are simulation EDA tools headed in the AI era?” with three conclusions:

Technically, AI will not replace physics solvers, but will change how solvers are invoked and verified.

From a product perspective, an Agent is not simply a conversational interface, but an engineering system jointly built from tools, data, verification, and governance.

For Julin Technology, the company will first focus on high-value, verifiable SI closed loops, and then gradually expand toward multi-tool and cross-physical-domain collaboration.

For simulation EDA, the changes brought by AI will ultimately be reflected in engineering results: Can the workflow deliver verifiable conclusions faster and more reliably?

Future competition will also revolve around this ability to continuously deliver such results.

Julin Technology will continue to advance the integration of high-precision simulation tools, engineering workflow automation, and trusted Agents, gradually validating each layer of capability in real-world design scenarios and providing chip design teams with more controllable and traceable simulation workflows.