Can Models Created With AI Be Trusted?

EDA models that are created using AI need to pass more stringent quality and cost benefit analysis compared to many AI applications in the broader industry. Money is hanging on the line if AI gets it wrong, and all the associated costs must be factored into the equation.

Models are some of the most expensive things a development team can create, and it is important to understand the value that will be derived from them. That value may come in several ways, such as being able to perform analysis faster or earlier in the development flow. The cost side has to be carefully considered, as well. Where does the data come from? how reliable is it? How much time or effort is involved in the creation of the model? How will it be verified? And finally, someone has to ascertain if the model is fit for purpose. Does it have the necessary accuracy or fidelity to provide the expected guidance? Will there be a way to validate that the model has resulted in good decisions being made? And how far is that detection point away from the decision point?

It should thus be no surprise that teams introduce a new model only when absolutely necessary. But in an era of shift left, and a growing need for system-level analysis, there is demand to be able to create models from prior designs, works in progress, or to enable feedback loops that can help automate aspects of the flow.

There is a plethora of ways that models, execution engines, and humans interact with each other, and there are already examples where AI is being used to replace or enhance each of those pieces. Among them:

• Creating an optimized model executed with a regular solver;
• Creating an optimized solver that executes a regular model;
• Creating a model from requirements, and
• Using a model of the process to replace or enhance the human in a feedback loop.

There are likely additional use models being used in the industry, and there certainly will be more in the future. There is also not one type of AI that is beneficial. “Keep in mind that many approaches do not need extensive datasets,” says Carlos Morales, vice president of AI at Ambiq. “Consider reinforcement learning, which effectively relies on data synthesis, transfer learning, and few-shot/no-shot learning. These approaches reduce the need for massive datasets.”

AI is not a panacea, and without fully understanding what is going on, there could be surprises. “It doesn’t matter how good the generated results are,” says Dan Yu, solution manager of AI ML for design verification team at Siemens EDA. “You still need some classic rule-based system to double check the generated result. AI, in the current form, will not immediately change or replace human workers directly. It is more like an assistant. There is always a mismatch, and machine learning by itself always can make mistakes. We will never be 100% certain of the results.”

One prevalent use case is to trade off accuracy for performance. “When you want better performance, or faster insights, you will have lost on accuracy,” says Mazen El Hout, senior product marketing manager at Ansys. “The loss of accuracy is very small, but the gains you have in terms of speed are very big. This would work for use cases when you don’t need high fidelity results. You would not use it for final validation. You’re at the design exploration phase, where the first priority is speed. A small loss of accuracy compared to traditional simulation is not a real problem. However, you can rely on it to get your needed insights.”

Having those checks and balances is important. “You want to start earlier because you need to start understanding what can be achieved,” says Marc Serughetti, senior director of product line management for Synopsys. “What are the challenges I may face? We know that building those systems of systems becomes more expensive, and we know that the further in the development process you find a problem, the more costly it is. You need to do things earlier and earlier, but it can only be done earlier with more abstraction because there are things you don’t know.”

Generated models are finding application in many areas. “Reduced-order models (ROMs) use AI to create fast-acting models of simulation data, helping create digital twins for electronics manufacturing, and improving yields in manufacturing,” says Srinivasa Mohan, distinguished engineer at Ansys. “We also can look at the impact of manufacturing on performance and optimize manufacturing processes to reduce variations. Besides that, ROMs also are used to accelerate design process optimization in IC chips, look at the impact of power profiles on thermal performance, and reduce the need for throttling.”

Understanding those models is essential to understanding how to use them safely. “AI is an automated way to discover useful rules and relationships hidden in data,” says Ambiq’s Morales. “In order for synthetic data to be useful, the creator of that data must express those rules. If the simulation used to produce the data doesn’t have those hidden rules built into it, the AI training process won’t find them. Anecdotally, we use simulation to augment our real-world ECG datasets. We found that in some cases, the model we trained with that synthetic data was seeing a ‘heartbeat’ signal in random noise, an effect of ‘hallucinating’ something that wasn’t there. Once we found this behavior, we had to add rules to our simulator to include that special case. In most domains, there are thousands or millions of rules, so it is hard to simulate everything.”

Physical world
Many types of models are used within the semiconductor industry, ranging from functional and verification models to physical models such as thermal. “Data synthesis is used effectively in situations where the data represents something physical, like a street or an airplane,” says Morales. “This is because physics is well understood, and we have decades of R&D creating simulations of the real world.”

There is always a pairing of a model and an execution engine. “You typically have a model, which is a virtual representation of a system based on physics,” says Ansys’s El Hout. “Then you use simulation to evaluate the performance, the stress, the thermal, or the displacement. In the semiconductor world, that might be the thermal behavior of a circuit board. What we’re doing with AI is using many previous simulation results on many different variations of different models — say, different variations of circuit boards — and then we are creating an AI based model of the solver. We are replacing the physics-based solver with an AI-based data-driven solver. With this newly created solver, we can predict performance of a new model very fast with reliable accuracy.”

But there are limitations. “Models trained via non-supervised learning (such as LLMs) are not amenable to this approach,” warns Morales.

Balancing attributes
The creation of a model is a balancing act between the time it takes to generate the model, the accuracy it provides, and its speed of execution. “Bringing technology that automates the creation of models is absolutely critical,” says Synopsys’ Serughetti. “The industry has to be careful about how much information you bring into your model. We all know that the more information you put in a model, the more accurate it becomes. But now you have issues with simulation performance.”

“Customers want to start with rough order of magnitude models, and then turn the screws up on the fidelity on those models,” says Shawn Carpenter, program director at Ansys. “But that comes at a cost. How long does it take to compute those models? In the context of the mission, you don’t want to embed a finite element simulation of an antenna into a flight simulation because you’re going to incur way too much time penalty. You have to tune the abstraction of the models. How do I get models — usually reduced-order models — that represent parts of the sub-system that are interesting to me, but which deliver a fidelity to me that makes my mission simulation accurate or my system-of-systems simulation accurate?”

In some cases, the lack of data can lead to problems with the model. “Coverage can be thought of as what percentage of the possible states have been synthesized,” says Morales. “Even for small domains, such as keeping a robot from falling, the number of states approaches infinity. This phenomenon is an obstacle to deploying AI models where synthetic data is a significant percentage of the training data. Still, to a lesser extent, it also applies to any model.”

Throughout the flow, different accuracies may be required. “At the beginning, you need high fidelity to train your AI,” says El Hout. “At the end of the workflow, you again need high-fidelity simulation to validate your design. AI-based prediction is only an acceleration for specific use cases in a specific design phase. When generating an AI model, it is important to know how long the user is willing to wait. If the user is willing to wait for two days for the model, then it will be more accurate, but if they want the model in two hours, it will be less accurate. The training time is inversely proportional to the accuracy and the number of data samples you have. The more data you have, the better the accuracy. The more time you’re willing to wait, the more accurate.”

Hallucinations
Not all AI-generated models are equally trustworthy. “In a large language model, people say it is not a prediction mistake, it is a hallucination,” says Siemens’ Yu. “But it is the same thing. We have learned that you can never trust AI in the current form for some tasks, and over the long run the strategy is that AI will complement our existing verification tools. It will make existing tools more powerful, more accurate, make them run faster because we don’t have to search a big problem space. But ultimately, AI still needs somebody sitting next to it to examine the result. We, as humans, understand our needs, which AI cannot immediately learn. AI does not have that knowledge today.”

But not all AI models will hallucinate. “If you want to recognize a cat, you use AI and train it with lots of picture of cats, and then it will recognize a cat,” says El Hout. “In this use case, it is very difficult to write a program that would be able to do this function. That’s why we use AI. But when we are looking at physical problems, we are already solving this without any issue. We do it very well, with high accuracy and fidelity, using simulation. The aim here is to accelerate, to make the prediction faster while accepting a loss of accuracy. It’s not because we don’t know how to do it. It’s because we want to go faster.”

Whenever AI can get it wrong, it is important to have the right checks in place. “When used in the refinement process, you need to provide the ability to come back and say, ‘I’m seeing a discrepancy,'” says Serughetti. “How do I address this discrepancy from that point in time? Or do I have the ability to come back to the previous step and review that discrepancy? So it’s a forward and backward flow that needs to take place in that model development, and how the models are evolving.”

Helping debug
More than half of development resources are used in verification. And within verification, debug consumes a huge chunk of time. If ever there was an application that could use help from AI, this would be it. “The data and modality of data is much more complex than other parts of the development flow,” says Yu. “We have code, we have verification results, we have waveforms, we have documentation, and we have specification. The complexity of functional verification is really high. We are thinking about a holistic approach to deploy that kind of data platform to help our customer gather their data, and to gain additional insight previously unavailable.”

Some of this is already happening. “We did an interesting test for waveform anomaly detection,” says Adam Arnesen, chief systems engineer for NI’s Test & Measurement Group within Emerson. “We took hundreds of billions of waveforms, time series information that was captured from buses and analog signal from a bunch of chips. We told it to identify events, the characteristics of those waveforms, and when they happened. We built a profile based of that information so we could ask what led to a particular issue with a register. Maybe it was a digital error somewhere, maybe there was a code error somewhere, maybe it was a transistor that was poorly printed and out of spec. You need a very large amount of information to figure that out.”

There are many places to look for data that can assist in pointing the user in the right direction. “You start by looking at the regression results, simulation results, and think about which bug is causing a specific failure,” says Yu. “By looking at all the available information, I can aggregate the results together and tell you that those failures are all talking about the same thing. They are pointing to the same design bug, and you need to look in that direction. This is the first step — to narrow down the search space. We also use machine learning to look at the historical commits and see which code has historically been verified many times, which we have more confidence in. Other code may have been committed by somebody over the weekend just to meet a deadline.”

You have to feed AI the right set of data, and you have to ask the right questions. “You have to learn to ask the right questions at the right level of breadth,” says NI’s Arnesen. “This can help you narrow down the state space that the human has to think about. It will identify the things that may be anomalies, and you don’t have to scroll through endless piles of data trying to find something that’s out of your comfort zone.”

Conclusion
The chip industry has barely scratched the surface about what may be possible with AI. While it is easy to get carried away with generative AI and ask when we will have the ability to design hardware from a loosely constructed, ambiguous specification, few will get any real value from that. Making the development of complex systems better, faster, and cheaper will be the real benefit, although it is unclear if all of the uses of AI being developed today will provide the claimed results.

“AI developers start from identifying some neat data and look at how they can create a model,” says Morales. “Instead, we want to identify an application and look at what data we need to make that possible.”