AslanX Flow in Action: Training & Deploying a CFD Surrogate Model | Deep Dive — Mason Seo
Mason Seo, CTO of Narnia Labs, demonstrates AslanX Flow, a platform enabling engineers to train, evaluate, and deploy surrogate models for 3D field prediction (e.g., CFD aerodynamics) without a dedicated machine learning engineer. The deep dive covers building a training workflow from raw engineering data, comparing model architectures like PointNet, Transolver, and Linear Neural Operator, and deploying the winning model as a no-code inference application. A key insight shared is the importance of evaluating models on specific engineering data, as public leaderboard performance doesn't always translate, alongside a preview of AI-driven workflow configuration.
This deep dive features Mason Seo, CTO of Narnia Labs, demonstrating AslanX Flow, a platform designed to connect raw engineering data to AI models without requiring a machine learning engineer. The core problem addressed is the computational cost of traditional engineering simulations (e.g., CFD runs taking hours per shape) and how surrogate models can provide predictions in seconds by sacrificing some accuracy for speed.
The Challenge of Surrogate Model Selection
Mason highlights a critical issue with publicly available leaderboards for surrogate models: the top-performing model can vary significantly based on the number of training runs and the specific dataset. A model ranking first on a public leaderboard might not be the best for an engineering team’s specific data. The presentation emphasizes the importance of measuring model performance on your own data rather than relying solely on generalized leaderboards.
Demonstration Data and Setup
The live demonstration uses a dataset of 64 solved CFD runs for vehicle aerodynamics. Each run includes a unique vehicle shape and its corresponding surface pressure. This dataset is split into 44 for training, 12 for validation, and 8 for testing, representing a typical, relatively small dataset that most engineering teams possess. The ultimate goal is to predict pressure on a vehicle shape the model has never seen during training.
AslanX Flow: Building and Training Workflows
AslanX Flow is presented as a node-graph interface for building, training, evaluating, and deploying AI models. The demonstration walks through creating a training workflow from scratch:
- Data Loading: Raw data (VTK files for solved results and CAD files for shapes) is loaded from storage.
- CAE File Inspection: The
inspect CAE filetask allows users to understand the structure of their simulation results, identifying key fields like ‘P’ (pressure) stored on nodes. - Pre-processing: Models typically require point clouds extracted from CAD surfaces. A task transforms the surface into a point cloud, normalizes the target field (e.g., pressure), and prepares it for model input.
- Data Splitting: The dataset is split into training, validation, and test sets (demonstrated with a 7:2:1 ratio).
- Model Training: Three different model architectures (PointNet, Transolver++, and Linear Neural Operator) are trained side-by-side for comparison. Resources (GPU, CPU, memory) are assigned to optimize training time. The platform allows monitoring training metrics like loss, relative L2 error, and R-squared.
- Model Evaluation: After training, models are evaluated on unseen test data. The demonstration shows how to compare R-squared values and relative L2 errors. In this example, the Linear Neural Operator (R-squared 0.94, relative L2 error < 20%) significantly outperformed PointNet (R-squared 0.63, relative L2 error ~50%).
AslanX Flow: Inference and Deployment
Once a suitable model is trained, the next steps involve creating an inference workflow and deploying the model:
- Inference Workflow Creation: A new workflow is created to load CAD files (STL) and use the trained model for prediction. Crucially, the
process info(mean, standard deviation, bounding box) from the training pre-processing step must be reused to ensure correct denormalization during inference. - Model Promotion: The best-performing trained model is promoted to a central model registry, making it available for use across different projects and workflows.
- Normal Convention: A common issue with CAD files is the direction of surface normals. AslanX Flow provides an option to ‘flip’ the normal convention to align with typical CFD results, ensuring correct predictions.
- Deployment to AslanX Studio: The
create deploymenttool in AslanX Flow allows transforming an inference workflow into a simplified, no-code inference application in AslanX Studio. This enables non-expert users to upload input files and receive predictions without interacting with the complex workflow graph.
AI-Assisted Workflow Creation
The session concludes with a demonstration of AslanX’s MCP (Model Creation Protocol) integration, showcasing an AI assistant (Claude) autonomously building and configuring a complete workflow from a single natural language prompt. This highlights the platform’s potential for further automation, allowing engineers to simply describe their needs, and the AI will configure the necessary tasks and settings within AslanX Flow.