A Unified AI–Quantum–Physics Framework for Autonomous Alloy Discovery and Multi-Property Materials Design
Akash Kanji
Department of Metallurgical and Materials Engineering, Jadavpur University
Visiting Researcher, Washington State University
The Alloy_Design project — codenamed Project TitanForge — aims to revolutionize alloy discovery by integrating Artificial Intelligence, Physics-Informed Modeling, and Quantum Computation into a single autonomous platform.
Our mission is to build a Universal Materials Intelligence System capable of learning from atomic structures, generating novel compositions, and predicting mechanical, thermal, electronic, and chemical properties — all while remaining physically interpretable and experimentally grounded.
We envision TitanForge as the next-generation digital forge for materials — where atoms, physics, and computation converge to accelerate innovation in steels, superalloys, and energy materials.
The repository is organized into four major research streams (A–D). Each stream contains specific modules (A1–D4) with code, datasets, and models.
"Teaching AI the language of atoms."
Focus: Large-scale representation learning and predictive modeling of materials properties.
- A1_ElementEmbeddings/ – Elemental and atomic vector representations using physics-informed descriptors and self-supervised learning.
- A2_GNN_MultiProperty/ – Multi-task Graph Neural Network (GNN) model predicting multiple materials properties (E, σ, κ, Eg).
- A3_TransformerGNN/ – Transformer-GNN hybrid capturing long-range interactions and crystal symmetry.
- A4_MatFold_Benchmark/ – Open benchmark for materials ML with standardized splits and cross-validation protocols.
Goal: Develop a foundation-scale model (MatFM) trained on 1M+ crystal structures to predict mechanical and electronic properties simultaneously.
"Designing materials backwards — from desired properties to atomic structure."
Focus: Generative AI for alloy and microstructure design using Diffusion, GANs, and inverse modeling.
- B1_DiffusionCrystalGen/ – Diffusion-based model for property-conditioned crystal structure generation.
- B2_InverseDesign/ – Neural inverse mapping: target property → optimal composition/structure.
- B3_StabilityAwareGen/ – Phase diagram–guided generator ensuring thermodynamic feasibility.
- B4_PolymerOrganicGen/ – Extending inverse design to polymeric and organic material systems.
Goal: Build a conditional generative engine that autonomously proposes stable, high-performance alloys and composites.
"Embedding the laws of physics and the power of quantum computation into learning."
Focus: Combining numerical physics, differential modeling, and quantum-enhanced algorithms.
- C1_PINN_PhaseField/ – Physics-Informed Neural Network for diffusion, segregation, and phase-field simulations.
- C2_DFT_Surrogate/ – Multi-fidelity surrogate model for Density Functional Theory (DFT) with uncertainty quantification (UQ).
- C3_QML_Superposition/ – Quantum ML module using variational quantum eigensolvers (VQE) for low-energy state search and superposition sampling.
- C4_MultiFidelity/ – Hierarchical surrogate learning linking empirical → DFT → experimental fidelity levels.
Goal: Achieve quantum-accelerated design through QML–PINN coupling for alloy phase and defect prediction.
"Closing the loop between AI, experiment, and reality."
Focus: Federated learning, multimodal data fusion, and closed-loop active learning.
- D1_MultimodalDataset/ – Curated datasets linking structure, XRD, spectra, micrographs, and process parameters.
- D2_ActiveLearning/ – Reinforcement and uncertainty-based experiment selection for robotic thin-film synthesis.
- D3_FederatedLearning/ – Privacy-preserving multi-lab data training for cross-institutional model robustness.
- D4_Explainability_Robustness/ – Counterfactual and out-of-distribution (OOD) detection tools for explainable materials AI.
Goal: Build an autonomous materials discovery loop — model suggests → experiment tests → model refines.