Industrial Engineering #5 - Pouria Mistani, Synopsys
Guest: Pouria Mistani | Senior Staff AI Engineer, Synopsys
One of the most exciting opportunities for AI lies in its ability to tackle computationally intensive problems. How can we harness these advanced systems to drive breakthroughs in fields like drug discovery and chip design? Not only is this month’s guest at the forefront of AI innovation, but he has seen firsthand the use cases for applying such advanced technologies to problems across. Pouria Mistani, a founding engineer on the GenAI Solutions team at Synopsys as a Senior Staff Engineer, spent his career building and implementing AI models for cutting-edge research in a wide variety of applications. Synopsys (NASDAQ: SNPS) is one of the foundational companies in the AI space, providing solutions across the design, verification, and manufacturing ecosystem for ‘silicon to systems’ solutions across a broad range of vertical markets. Last year, the company achieved over $5 billion in top-line revenue, representing a nearly 15% YoY increase, and today commands an almost $75 billion market cap. My discussion with Pouria covered everything from the critical importance of partial differential equations (PDEs) to the tools like JAX, HuggingFace, and MLFlow powering his workflow to his view on the growing importance of neural operator-based simulators. Let’s dive in!
Pouria received his PhD in Mechanical Engineering with an emphasis in Computational Science and Engineering (CSE) from University of California Santa Barbara in 2020, and his MS degree in Physics with an emphasis in computational astrophysics from University of California Riverside in 2014. Prior to that, he received two B.S. degrees in Aerospace Engineering and Physics from Sharif University of Technology in Iran in 2013. Pouria did his postdoctoral studies at Merck Research Lab, where he developed computational models for high concentration biotherapeutic formulations.
Pouria’s research interest is focused on computational modeling and simulation of multiscale phenomena. In his career he has applied advanced numerical techniques to develop computational models and HPC simulation software for several complex physical systems including galaxy clusters, cell aggregates, atomic islands, instabilities of high concentration protein formulations, parametrization of molecular dynamics force fields, and using AI for accelerating molecular dynamics simulations.
In his personal time, he enjoys reading books, listening to podcasts and music, stargazing, and going for long walks.
Your experience in AI has covered multiple industry sectors. How did you move from drug discovery to chip design?
I came to the US to pursue a PhD in Physics, but my work on the Illustris Simulation Suite made me realize the transformative role of high-fidelity, large-scale simulation software in uncovering new insights about nature. Recognizing computation as a foundational pillar of scientific inquiry alongside traditional experimentation, I decided to complete a PhD in Computational Science and Engineering. My research focused on partial differential equations (PDEs) for physical systems with evolving interfaces, such as those encountered in materials science and biology, with an emphasis on developing multiscale simulation engines to address these challenges. This led me to explore the applications of these techniques in industry use cases, and the first one I considered was biotherapeutics for drug discovery. Between 2020 to 2024, I was working on AI applications for drug discovery but I eventually transitioned to chip design.
I think it's too early for drug discovery to be revolutionized by AI because the key problem is the significant lack of data and the significant sensitivity of biostructures to their environment; it’s not just an issue of computational resources or training a new model. These applications are still at least five years away from widespread adoption and efficacy. For drug discovery, the available data set is very limited and can only be obtained by running scientific experiments. Because this is so expensive and highly specialized, there are only about 227,000 experimentally determined 3D biostructures available to date. Even these structures are not readily suitable for developing AI models for drug design because to obtain biostructures, researchers often run structural mapping experiments on proteins, which typically focus on stable states. However, biologically relevant proteins, including biotherapeutics of interest in drug discovery, are often metastable and contain intrinsically disordered regions that lack a stable conformation. Consequently, structural data banks tend to be biased toward more stable protein structures. This bias is reflected in designs produced by AI-based structure prediction algorithms, which predominantly model stable conformations. For example, proteins in the human body often exhibit metastability—during a fever, even a slight increase in body temperature can lead to structural conformations shifting and aggregating, altering their functions. This dynamic nature makes biologically relevant proteins more challenging to model and predict effectively, particularly in drug discovery contexts.
Biological structures adapt to their environmental conditions, even small energy differences captured by quantum calculations can lead to significant conformational changes in the protein structure, which in turn alters its function. This sensitivity is crucial for the precise design of drugs, and nature is not forgiving when it comes to this level of complexity. On the other hand, chip design is a comparatively easier problem. Unlike drug discovery, where quantum mechanical calculations are essential to understanding the behavior of molecules, chip design involves a static system with known components. Every part is modeled with well-established equations, and while the problem may be high-dimensional, the complexity lies in the scale. As chips become larger and more complex, the primary challenge becomes modeling the entire chip floor plan, which can be computationally intensive.
Although the problems in drug design and chip design are distinct, there is a mathematical similarity between the "structure prediction" problem in drug discovery with the "placement and routing" problem in chip design. Both involve optimizing the arrangement of components—whether it’s placing atoms or transistor blocks in the most effective configuration. While the details of the two problems differ, this common task of arranging elements offers an opportunity to transfer methodologies between protein design and chip design, and vice versa. Insights and techniques developed in one field could potentially spill over to enhance the other.
What is the biggest technical bottleneck you've found when applying AI to such computationally intensive problems? How have recent AI advancements either helped or hindered these questions of scale?
A major bottleneck lies in the scalability of algorithms to handle multi-scale problems inherent in domains like drug discovery and chip design. For example, in molecular simulations or PDE-constrained optimizations for chip design, you often encounter a mix of fine-grained dynamics and long-range interactions that require enormous computational resources. Bridging this gap between resolution and efficiency remains a challenge.
Recent advancements, like the advent of transformer-based architectures and diffusion models, have helped approximate some of these systems more effectively, especially in learning latent representations of high-dimensional spaces. However, these models often demand substantial computational resources, which can exacerbate accessibility challenges for smaller teams or startups. While tools like JAX (from Google) have made it easier to prototype differentiable solvers, balancing accuracy with efficiency remains an unsolved problem.
With JAX, you don’t need large teams to build performance-scale parallelizable AI models anymore and one person can write code that can easily be partitioned across a complex system on most AI accelerators. In my previous roles, there was always an internal debate between teams to use PyTorch or JAX. Often, companies choose PyTorch because of their familiarity with the system. When using PyTorch, developers can achieve good performance on NVIDIA GPUs because the underlying layers are optimized by hundreds of engineers at NVIDIA but if you need any custom design components or use other hardware accelerators, you often don’t readily have any of the fully fused kernels and highly-optimized layers. When you try putting the pieces together in the model, you lose performance. However, for JAX, this is managed by the developer at the API level, and the state-of-the-art compiler in JAX can efficiently optimize it for various hardware accelerators. When you run tests to benchmark the two systems, JAX is always the clear winner. Nowadays, leading AI companies, such as DeepMind and AWS, already use JAX because achieving the required scale and performance with PyTorch would necessitate extensive manual low-level optimizations.
There are numerous projects that stalled after developers picked the wrong technology because they could solve a small test problem quickly, but the technology stack fails to support the scale of the actual problem. But by that point, they’ve already invested so many resources into using that technology stack so it’s too hard to go back. In my opinion, JAX should be the de facto framework across AI enterprises.
How have tools like JAX impacted your engineering workflow?
Differentiable programming frameworks like JAX have been transformative in my workflow, particularly for AI in scientific applications. JAX’s novel functional programming paradigm enables clean, composable code, making it easier to experiment with complex architectures while achieving state-of-the-art performance. Its seamless support for automatic differentiation, GPU/TPU acceleration, and parallelization has made scaling models—from small prototypes to large-scale simulations—straightforward. This is especially critical when working on computationally intensive tasks like neural PDE solvers or stochastic modeling, where both precision and efficiency are essential.
In addition, advancements in the open-source ecosystem—such as tools from HuggingFace—have played a pivotal role in accelerating AI development. HuggingFace’s libraries have democratized access to state-of-the-art models and methods, enabling rapid experimentation and benchmarking.
Complementing these tools, code observability platforms like MLFlow have been instrumental in ensuring reproducibility and monitoring experiments. In high-stakes environments, where validation and debugging are critical, MLFlow’s capabilities make it easy to track model performance, manage experiments, and maintain a clear audit trail for iterative development.
By combining JAX’s scalability, open-source innovation, and MLFlow’s observability, I’ve been able to build robust, high-performance solutions tailored to scientific and engineering challenges.
What is one tool you are actively evaluating? What factors do you typically consider before adding a new system to your workflow?
I'm currently evaluating Ray for distributed computing, particularly in multiagent reinforcement learning workflows. With my recent focus on multiagent systems, efficiently scaling experiments and managing heterogeneous workloads is increasingly important. Ray's flexibility in handling distributed systems while integrating well with Python-based ML stacks makes it a promising option.
When considering new tools, I prioritize:
Interoperability: Does it integrate well with existing systems?
Scalability: Can it handle increasing workloads or parallelism without significant overhead?
Community and Documentation: Are there sufficient resources for troubleshooting and collaboration?
Cost-Effectiveness: Does it justify the ROI, especially in resource-intensive projects?
When I consider any new technologies, my biggest question is always the number of blockers to adopt and use the system. Any new tool that I evaluate has to be open source. I don’t like systems that are third-party-owned since that means I have no control and when developing new models where the specs and requirements are still in flux, it’s really important to maintain ownership. With Ray, for example, the architecture is already quite modular, and their codebases, white papers, and research blogs are all publicly shared, which helps understand the underlying code. Companies develop trust by sharing their architecture and roadmap, and showcasing their GitHub pages with code and community contributions.
Enterprises often prefer to use private LLM instances due to security concerns. Depending on the deployment type, cloud providers can share some responsibility (such as network controls or physical host security for a SaaS deployment with Microsoft), which helps build customer trust. Companies running their models locally sometimes lack the resources to scale but the token rate for local models is usually low. Cost isn’t as much of an issue if the company offers robust support.
What is one piece of advice about technical architecture choices that you would share with engineers who are looking to build AI solutions for deep tech applications?
Design architectures that embrace modularity and adaptability from the outset. Deep tech applications often evolve as new insights emerge, and rigid systems can quickly become a liability. For example, when developing AI models for PDE solvers, using modular pipelines—where ML components like encoders and decoders can be swapped—allows teams to iterate efficiently without overhauling the entire system.
Equally important is adopting containerization technologies like Docker for environment management. AI projects are particularly prone to complex dependencies, large software packages, and compatibility challenges across different systems. Docker provides a robust, standardized way to package, deploy, and scale AI solutions, ensuring that models run consistently across development, testing, and production environments.
I’ve seen firsthand how companies struggle in their software development lifecycle when they forgo containerization—often due to security concerns—which later proves to be a costly decision. Properly configured Docker containers, paired with enterprise-level security practices, resolve these concerns while enabling reproducibility, scalability, and faster onboarding for development teams.
Finally, prioritize domain-informed models over purely data-driven approaches. Integrating physical constraints or expert priors enhances model robustness and interpretability, especially in deep tech applications where reliability and scientific accuracy are paramount.
Within the AI infrastructure space, what is one specific new technology trend you're watching?
I'm particularly excited about neural operator-based simulation tools and their potential to accelerate large-scale simulations, enabling spatiotemporal scales that were previously out of reach. These tools address critical bottlenecks in computationally intensive fields like drug discovery and chip design, where traditional methods struggle to scale efficiently.
For example, in drug discovery, I’ve worked on developing neural operator-based simulations for high-concentration protein formulations. These approaches allow us to reach the extended time scales required for designing better biotherapeutics—something that was infeasible with traditional solvers. Similarly, in chip design, spatiotemporal scale limitations arise when coupling electronic design automation (EDA) tools with multiphysics constraints. Neural operator-based methods can alleviate these challenges by providing efficient and scalable approximations of complex mathematical models in the design process.
Looking ahead, my ultimate goal is to enable the creation of AI design agents that leverage these AI-accelerated simulation tools across scientific domains and industries. Advances in AI algorithms, combined with modern AI frameworks (i.e., JAX) and compute platforms, now make it possible to integrate neural operators into autonomous systems. These physics-informed agents would be capable of rapidly exploring design spaces, optimizing solutions, and driving innovation across deep tech applications without the computational bottlenecks of traditional approaches.
Our role as developers and engineers will also change as these tools improve. Neural operators can only act as accelerators so over time, design agents will be able to work more effectively as they can consider electromagnetic, thermal, and other physical constraints and decide how to adjust. The decision-making and reasoning parts of interpreting the outcomes and modifying the design parameters will be much harder to automate. Ideally, you want to make these models so good they can design a new chip, and the role of humans will be to provide the agents with goals, expected performance, and constraints.
This series focuses on navigating technical software decisions within industrial companies. From optimizing infrastructure choices and leveraging DevOps best practices to harnessing the power of cloud technologies and improving data workflows, our guests will highlight how they've considered these decisions and implemented new solutions across their organization. If you're similarly excited about leveraging technology to empower our national industrial base and/or building solutions focused on this category, please reach out. My email is ananya@schematicventures.com – I’d love to connect!