Get your free personalized podcast brief
We scan new podcasts and send you the top 5 insights daily.
While semiconductor architectures and advanced hardware schematics are guarded behind strict corporate firewalls, the underlying physical laws governing heat, energy balance, and electromagnetism are universal. Hardik Kabaria of Vinci explains that foundation models trained on fundamental physics equations can simulate and validate complex multi-physics chip designs out-of-the-box without requiring access to proprietary customer training datasets.
Related Insights
The biggest obstacle to fully AI-driven hardware design is the absence of a large, public training dataset. Unlike software code, circuit board designs are proprietary and siloed within companies like Apple and SpaceX. Until this data is generated or aggregated, model capability will be constrained, regardless of architectural breakthroughs.
Current AI models for science are narrow surrogates for specific tasks. The grand vision is to build a foundation model for physics that understands a wide range of coupled, multi-physics phenomena. This single model could be used for simulation, inverse design, and control across many scientific and engineering domains.
Arena Physica has built the first foundation model for electromagnetism. Since no public dataset exists, they created a "data factory" to generate training tokens of materials and electromagnetic fields, allowing the model to design novel RF hardware from a text prompt.
Startups and major labs are focusing on "world models," which simulate physical reality, cause, and effect. This is seen as the necessary step beyond text-based LLMs to create agents that can truly understand and interact with the physical world, a key step towards AGI.
The next leap for hardware—AI generating complex 3D CAD designs—is blocked by a data bottleneck. CAD files are a company's most valuable IP, so firms won't share them to train models. The solution may lie in on-premise models or starting with the hobbyist community.
Instead of training models on scarce circuit board data, Diode Computers built a compiler that makes hardware design look like a Python program. This allows powerful language models, which are expert coders, to design physical hardware by leveraging their existing capabilities, bypassing the data bottleneck.
Large Language Models are limited because they lack an understanding of the physical world. The next evolution is 'World Models'—AI trained on real-world sensory data to understand physics, space, and context. This is the foundational technology required to unlock physical AI like advanced robotics.
While traditional hybrid models still require solving differential equations during use, Physics-Informed Neural Networks (PINNs) learn the final outcome directly. They are trained to obey physical laws but don't need the equations for inference. This provides a significant speed advantage for computationally expensive systems like chromatography, making them more suitable for real-time applications.
For physical design, simulation shouldn't just be a final verification step. Instead, it should be a tool used during model training to build the AI's intuition or "taste." This allows the model to generate high-quality designs quickly at inference time, mirroring how expert human engineers develop their skills.
The ability to rapidly simulate complex hardware is the single biggest unlock for deep tech investing. Where it once took years to run simulations for a new reactor, it can now be done in hours. This compresses the hardware development cycle, making it fast enough to fit within venture capital timelines and expectations.
