| --- |
| license: mit |
| tags: |
| - aerospace |
| - physics-informed-neural-network |
| - pinn |
| - pytorch |
| - surrogate-model |
| - alrithm |
| library_name: pytorch |
| --- |
| |
| # 🚀 ALRITHM-Astro |
|
|
| **ALRITHM-Astro** is an AI-powered Aerospace Engineering Copilot designed to simulate rocket flight dynamics and predict Maximum Dynamic Pressure (Max Q). It combines a custom-trained Physics-Informed Neural Network (PINN) with a natural language interface. |
|
|
| 🔗 **[Try the Live Demo Here](https://godela-web.vercel.app)** |
|
|
| ## 🏛️ Powered by Alrithm |
| This project is fundamentally built upon and inspired by the vision of **[Alrithm](https://alrithmapi.vercel.app)**. |
|
|
| - **Data Source:** The current model version is trained on a specific subset of the **`alr-aerospace-platform-0.2`** dataset (targeting the `max_q_kpa` quantity), fetched via the official Alrithm API. |
| - **Philosophy:** Alrithm provides rigorously verified, high-fidelity aerospace data. ALRITHM-Astro serves as a demonstration of how developers can leverage Alrithm's verified data streams to build real-time, physics-informed AI applications. |
|
|
| ## ✨ Features |
| - **Natural Language Processing:** Understands complex aerospace queries using Groq's LLMs (`openai/gpt-oss-120b`). |
| - **Physics-Informed Prediction:** Calculates Max Q with high precision using a custom PyTorch Neural Network. |
| - **Interactive 3D Visualization:** Renders the rocket's 3D trajectory in real-time using Plotly.js. |
| - **Full-Stack Architecture:** FastAPI backend (Render) + Vanilla JS frontend (Vercel). |
|
|
| ## 🧠 Model Details (Version 0.2 - Max Q Specialist) |
| - **Architecture:** Multi-Layer Perceptron (MLP) with 256 hidden units and SiLU activation. |
| - **Input Features:** 13 engineered physical features (e.g., Burnout Velocity, Thrust-to-Weight Ratio, Mass Ratio). |
| - **Output:** Maximum Dynamic Pressure (`max_q_kpa`). |
| - **Training:** 5-Fold Cross-Validation (Fine-tuned specifically on the `max_q_kpa` subset of the Alrithm dataset). |
|
|
| > ⚠️ **Current Limitation:** This `v0.2` model is a **Single-Task** surrogate model. It is currently specialized only for `max_q_kpa`. It has *not yet* been trained on the entire 4600+ dataset across all task types. |
|
|
| ## 🗺️ Roadmap (Next Steps) |
| - [x] **Phase 1:** Build the End-to-End Pipeline (Data -> Neural Network -> FastAPI -> Vercel). |
| - [x] **Phase 2:** Train a Single-Task Model for `max_q_kpa`. |
| - [ ] **Phase 3:** Implement **Multi-Task Learning (MTL)** using the full 4600+ `alr-aerospace-platform-0.2` dataset to predict all quantities (Apogee, Burnout Speed, etc.) simultaneously. |
|
|
| ## 🛠️ How to use this model |
|
|
| You can load this model directly using PyTorch: |
|
|
| ```python |
| import torch |
| import torch.nn as nn |
| |
| class PureMLPModel_Deep(nn.Module): |
| def __init__(self, in_dim=13, hidden=256): |
| super().__init__() |
| self.net = nn.Sequential( |
| nn.Linear(in_dim, hidden), nn.SiLU(), |
| nn.Linear(hidden, hidden), nn.SiLU(), |
| nn.Linear(hidden, hidden), nn.SiLU(), |
| nn.Linear(hidden, 1), |
| ) |
| def forward(self, x): |
| return self.net(x).squeeze(-1) |
| |
| # 1. Load the ALRITHM-Astro model weights |
| checkpoint = torch.load("alrithm_astro_model.pth", map_location="cpu") |
| |
| # 2. Initialize the architecture and load weights |
| model = PureMLPModel_Deep() |
| model.load_state_dict(checkpoint['model_state_dict']) |
| model.eval() |
| |
| # 3. Retrieve the scalers for data normalization |
| scaler_x = checkpoint['scaler_x'] |
| scaler_y = checkpoint['scaler_y'] |
| |
| print("✅ ALRITHM-Astro model loaded successfully!") |
