TheraJect’s Inference Materials Platform™ combine diverse generative proposal mechanisms with inference-time Pareto decision layers, physics-informed anisotropy priors, and SOC-DFT validation to efficiently discover REE magnet candidates.
At its core is TheraJect’s Pareto-Guided Inference Loop™, which enables inference-time trade-off optimization across performance, stability, cost, and manufacturability.
Architecture of pipeline
AI/ML Inference Engine
Our AI layer enables inference-time exploration and decision-making, not just offline model training. It integrates:
- Generative models (VAE, GAN, DDPM) for rapid candidate proposal
- Graph neural networks (GCN, GAT, MEGNet, M3GNet) for structure-aware prediction
- Transformer-based property predictors for multi-target estimation
- Materials-specific language models (MatBERT, Materials NLP) for composition-property inference
These models operate together in a Pareto-guided inference loop, enabling rapid exploration of vast chemical and structural spaces under real-world constraints.
Physics-Based Simulation(DFT)
We perform scalable first-principles simulations to validate and refine AI-generated candidates, including:
- Formation energy and thermodynamic stability
- Magnetic properties (Ms, MAE, Tc)
- Electronic structure and band gaps
- Crystal relaxation and structural optimization
Our workflows leverage GPAW, ASE, PyXtal-based structure enumeration, and ML-accelerated relaxations (M3GNet), with compatibility for VASP-style workflows when needed.
Multi-Objective Decision & Optimization
Candidate materials are evaluated using Pareto-based, multi-objective ranking, balancing:
- Thermodynamic stability
- Magnetic performance (Ms, MAE, Tc)
- Semiconductor or battery-specific metrics
- Cost, supply-chain criticality, and manufacturability
This approach enables transparent trade-off analysis and risk-aware down-selection.
Data Infrastructure
The platform integrates curated and synthetic datasets, including:
- Materials Project, OQMD, JARVIS, AFLOW
- Matminer-derived feature databases
- Synthetic enumerations of REE and transition-metal systems
- AI-generated datasets from generative models
All data streams are normalized and continuously refined through feedback from simulation and experiment.
Inference-Time Discovery Loop
- AI-based candidate generation
- ML-driven property prediction
- Pareto-guided multi-objective filtering
- Physics-based validation (DFT)
- Down-selection for experimental and manufacturing feasibility
This framework reduces weeks-to-months of exploratory work to days or hours, while improving hit-rate and interpretability.
TheraJect’s platform is built on open scientific methods.
Our contribution lies in integrating these methods responsibly for real-world materials discovery, under physical, numerical, and societal constraints