Industrial and Manufacturing Sector # Alsemy Accelerates Chip Modeling From Months to Minutes With Physics-Informed AI [Learn More](https://www.nvidia.com/en-us/industries/semiconductor.md) Alsemy ### Objective Founded in 2019 in South Korea, Alsemy set out to remove major bottlenecks in semiconductor innovation: the manual, equation-heavy process of building accurate device models for next-generation chips. Using physics-informed AI and large physics models in its Alsis and Alsphere platforms, Alsemy helps chipmakers replace year-long SPICE or T-CAD model development cycles with workflows measured in minutes or days while maintaining the accuracy required for advanced process nodes. To achieve this, the company turned to [NVIDIA RTX™ 3090](https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090-3090ti.md) GPUs, the [NVIDIA PhysicsNeMo™](https://developer.nvidia.com/physicsnemo) framework, and [NVIDIA® CUDA®](https://developer.nvidia.com/cuda/toolkit)-accelerated deep learning to train and deploy AI models that capture complex transistor behavior at scale. ### Customer Alsemy ### Partner AWS Google Cloud ### Topic Accelerated Computing Tools & Techniques ### Products * [NVIDIA CUDA](https://developer.nvidia.com/cuda) * [NVIDIA PhysicsNeMo](https://developer.nvidia.com/physicsnemo) * [NVIDIA RTX GPUs](https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus.md) ### Key Takeaways * Cut device modeling cycles from years of physics modeling and months of manual fitting to a few weeks of GPU training, plus one-second inference and 10-minute customer-specific tuning. * Reduced process development time and product time to market while lowering R&D costs, helping leading South Korean chipmakers keep aggressive roadmaps on track. * Enabled fabrication (fab) and R&D teams to reach expert-level model accuracy and serve more customers in parallel with the same core team. ## Reframing Semiconductor Modeling for the AI Era As chips pack more performance into less space, the underlying physics becomes harder to model accurately—and traditional workflows can’t keep up. Device modeling has depended on small groups of Ph.D.-level experts hand-tuning physics equations and parameters for each new technology node, a process that can take months to years for a single SPICE or T-CAD model. This slow, manual approach creates a critical bottleneck in design-technology co-optimization, delaying circuit design, verification, and ultimately time-to-market for new consumer electronics, data center chips, and other advanced devices. Alsemy saw that unless modeling became both faster and accessible to more engineers, chipmakers would struggle to meet advanced-node roadmap commitments. Alsemy Alsemy An AI-driven semiconductor manufacturing framework that continuously feeds real silicon data into a virtual lab to improve predictive accuracy and manufacturing productivity. ## Building Physics-Informed AI Models on NVIDIA RTX GPUs To break this bottleneck, Alsemy developed Alsis, an AI-driven device modeling platform that blends neural networks with explicit physical constraints, enabling physics-informed models that learn directly from I–V and C–V data while respecting underlying device behavior. In parallel, Alsemy built Alsphere for structural and process modeling, with a display-specific variant, DPS, co-developed with LG Display. The team trains large pretrained models, including a Retarget model and a C–V model, using [NVIDIA RTX 3090 GPUs](https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090-3090ti.md) in a hybrid infrastructure that combines on-premises systems with cloud experience from Amazon Web Services and Google Cloud Platform. Using NVIDIA PhysicsNeMo alongside CUDA-accelerated PyTorch, the Alsemy team is able to rapidly iterate through different model architectures during model development and reduce the cost of training large-scale, physics-informed models. The rapid ideation and experimentation using off-the shelf-modules of PhysicsNeMo has enabled the six-person (and growing) AI research and engineering team to operate as a focused AI modeling factory for semiconductor customers. Once trained, Alsis models are typically deployed for CPU-based inference in customer environments, meaning chipmakers benefit from GPU-accelerated training without needing to overhaul their production infrastructure. Partners can also train custom neural networks directly inside Alsis with their proprietary I–V datasets, giving them a controlled and secure way to encode their process know-how into high-accuracy device models. Across deployments, Alsis' pretrained and customer-specific models often run side by side to support circuit-level simulations and device performance evaluation in production flows. Alsemy ## Turning Expert-Only Modeling Into a Scalable AI Service Alsemy’s NVIDIA-powered, physics-informed AI workflow has transformed semiconductor device modeling from a slow, expert-only process into a fast, data-driven engine for R&D and production fabs. Where baseline physics-based models once took years to build and weeks to months of manual parameter fitting per device, Alsemy now delivers complete modeling solutions in just a few weeks of GPU-accelerated training, followed by one-second inference and roughly 10 minutes of customer-specific fine-tuning. This shift compresses time-to-model so dramatically that teams can iterate rapidly on new device architectures, enabling faster DTCO loops and more aggressive technology roadmaps across customers such as SK hynix, LG Display, and the National Nano Fab Center. Feedback from close collaboration with SK hynix indicates that model accuracy from Alsemy’s solution is comparable to that of a highly experienced process engineer, highlighting three core advantages: reducing R&D turnaround time with automation, avoiding human error that can impact product quality and performance, and enabling further AI-assisted R&D process improvement. At the National Nano Fab Center, R&D researchers and smaller companies can more easily access fab resources by first creating a digital twin of their process, significantly improving their chances of “one-shot” product success. These results are especially meaningful in an environment where, as Dr. Jun-Mo Yang, Principal Research Scientist at National Nano Fab Center, notes, “The importance of manufacturing AI and physical AI is widely recognized, [but] revolutionizing high-tech R&D systems remains extremely challenging” due to legacy processes and limited access to real industrial datasets. Alsphere is also being adopted as the standard process-prediction model for “BANDI,” the national semiconductor and display data platform NNFC is building with KISTI, opening the door for broader use across publicly funded R&D programs. For Alsemy’s customers, the impact shows up directly in business results. Process development times and product time‑to‑market shrink, R&D costs drop, and better model accuracy supports higher yield. Because Alsis can be quickly fine‑tuned with customer data and run on CPUs, fabs keep sensitive datasets on‑prem while still benefiting from NVIDIA GPU‑accelerated training. This approach gives Alsemy rare visibility into real manufacturing challenges and, using NVIDIA PhysicsNeMo as a foundation, enables scalable, physics‑informed AI models for semiconductors—setting the stage for a long‑term collaboration that advances both customer roadmaps and the broader PhysicsNeMo ecosystem. “PhysicsNeMo is a remarkably powerful tool for training AI on the complex physical phenomena of semiconductors. What impressed us most is how NVIDIA, through close collaboration with diverse industry partners and research institutions, is systematically translating real-world needs into reusable libraries. This ecosystem-driven approach is meaningfully accelerating the AI transformation of the semiconductor industry.” **Hyunbo Cho** CEO, Alsemy ## Paving the Way for Next-Generation Device Design Looking ahead, Alsemy plans to continue expanding its library of large physics-informed models and refining Alsis to support more device types, nodes, and customer-specific workflows. As process technologies advance and design complexity increases, the combination of NVIDIA GPUs, PhysicsNeMo, and Alsemy's domain expertise positions the company to become a central AI modeling partner for semiconductor companies that want to keep their roadmaps on track. With additional hiring for AI researchers and engineers and growing adoption among top-tier South Korean chipmakers, Alsemy is building toward a future where AI-driven modeling is a standard part of device development, helping the industry move faster while maintaining the precision it demands. Looking ahead, Alsemy plans to extend its AI physics leadership into agentic AI, using pretrained agents that absorb physics knowledge and autonomously perform prediction and optimization to transform the paradigm of semiconductor and display R&D process. Drive semiconductor breakthroughs with AI, digital twins, and accelerated computing. [Learn More](https://www.nvidia.com/en-us/industries/semiconductor.md) ## Related Customer Stories