An open-source research and development repository leveraging cutting-edge machine learning architectures—including Heterogeneous Graph Transformers (HGTs), Self-Supervised Earth Masked Autoencoders (EarthMAE), Constrained Optimizers, and Multi-Modal Foundation Models—to solve complex topology modeling, predictive, and capital allocation challenges in industrial sustainability and climate economics.
This repository is organized into four primary pillars, each addressing a critical data, architectural, or economic bottleneck in global industrial decarbonization and product lifecycle pipelines.
-
File Link:
GHG_Scope3_MultiTier_HGT.ipynb -
Deployment & Rendering:
- Objective: To accurately model non-homogeneous entity metadata and deep physical topologies inherent in global, multi-tiered supply chains under severe data-scarcity conditions.
-
Methodology: Implements a Heterogeneous Graph Transformer (HGT) architecture to capture the complex, multi-layered dependencies between upstream suppliers, refining facilities, logistics nodes, and end-products. Defined formally by the graph schema
$\mathcal{G}=(\mathcal{V}, \mathcal{E}, \mathcal{T}_v, \mathcal{T}_e)$ , the model leverages structural attention to propagate carbon intensity metrics across deep, unmapped supply networks. -
Key Achievements:
- Topology Resilience: Successfully mapped deep upstream supply chain layers, robustly handling sparse primary data and highly skewed entity distributions.
- Baseline Outperformance: Achieved a significant reduction in Mean Absolute Percentage Error (MAPE) for unmapped Tier-3+ supplier emissions compared to traditional, homogeneous Graph Convolutional Networks (GCNs).
- Attention Interpretability: Extracted learned edge-attention weights to isolate and identify high-variance carbon chokepoints across the simulated industrial network.
- File Link:
EarthMAE_Self_Supervised_Final.ipynb - Deployment & Rendering:
- Objective: To extract robust, high-fidelity representations from unlabelled remote sensing and geospatial data for scalable environmental monitoring and physical risk mitigation.
- Methodology: Leverages a Self-Supervised Vision Transformer (ViT) backbone with a masked autoencoder pre-training strategy. The model learns fundamental spatial-temporal patterns by reconstructing high-percentage masked imagery, creating highly transferable embeddings for downstream tasks like biomass estimation, land-use classification, and asset-level physical climate risk assessment.
- Key Achievements:
- High-Ratio Reconstruction: Demonstrated robust spatial reconstruction capabilities even when operating at a heavy 75% pixel masking ratio, proving deep contextual understanding of regional land-cover features.
- Data Efficiency: Downstream fine-tuning on highly limited labeled datasets achieved target convergence with substantially fewer labeled samples compared to a Vision Transformer trained completely from scratch.
- Transferability: Embeddings demonstrated exceptional zero-shot stability across distinct geographic biomes, minimizing domain-shift errors when transferring from training regions to unmapped test zones.
- File Link:
Dynamic_MACC.ipynb - Deployment & Rendering:
- Objective: To transform static, deterministic carbon abatement graphs into multi-period, path-optimized investment decision-support engines for multi-site operations.
- Methodology: Replaces traditional static MACC frameworks with a dynamic optimization engine. It models changing regulatory landscapes (e.g., carbon tax trajectories), capital expenditure depreciation, and technological learning curves over a rolling horizon, enabling industrial complexes to optimize their capital allocation paths under macroeconomic uncertainty.
- Key Achievements:
- Non-Convex Optimization Convergence: Successfully formulated and solved a multi-period, constrained optimization problem, effectively avoiding local minima to isolate global path-optimal capital allocation strategies.
- Capital Efficiency: The dynamic pathway identified an optimized deployment schedule that reduces the Net Present Cost (NPC) of industrial decarbonization over an extended 20-year horizon compared to static, cost-ranked baselines.
- Sensitivity Scaling: Built-in multi-variable sensitivity analysis allows real-time shifts in carbon tax trajectories and technology learning rates, updating the optimal investment sequence computationally in seconds.
-
File Link:
AI_LCA_fixed.ipynb -
Deployment & Rendering:
- Domain & Standards: Sustainable AI · Industrial Ecology · ISO 14040/14044 · GHG Protocol · EU CSRD (Digital Product Passport)
- Objective: To automate, predict, and generate highly granular Product Carbon Footprints (PCFs) and complete Life Cycle Assessments at scale, bypassing traditional consulting bottlenecks ($10K–$100K costs and multi-month turnarounds).
- Methodology: Implements CarbonAI v2.0, an end-to-end multi-modal foundation model framework trained on real-world public LCA registries (OpenEPD Building Transparency, EPA GHG Emission Factors Hub, DEFRA UK Conversion Factors, and USDA LCA Commons). The architecture fuses unstructured Bill-of-Materials (BOM), manufacturing process topologies, and regional grid data. It structurally integrates differentiable physics constraints to preserve mass/energy balances, proximal causal learning to uncover structural emission drivers, conditional diffusion for generative material synthesis, and constrained multi-objective Reinforcement Learning (RL) for design-space exploration.
-
Key Capabilities:
- Physics-Preserving Latent Space: Enforces physical laws (thermodynamics and material conservation) via custom differentiable loss penalties during backpropagation, avoiding hallucinatory carbon accounting.
- Generative Alternative Synthesis: Utilizes conditional diffusion to generate optimized product formulations and manufacturing adjustments that satisfy strict carbon reduction targets without compromising engineering specifications.
- Causal Attribution: Moves beyond simple correlations to isolate true causal vectors of environmental impact across complex multi-step industrial processes.
- Generative AI & Deep Learning: PyTorch, Vision Transformers (ViT), Conditional Diffusion Models, Multi-Objective Reinforcement Learning (RL)
- Graph Deep Learning: PyTorch Geometric (PyG), Heterogeneous Graph Transformer (HGT) architectures, Attention Mechanisms
- Mathematical Optimization & Causal Inference: SciPy Optimization, NumPy, Pandas, Proximal Causal Learning Frameworks, Non-convex and Constrained Optimization
- Interactivity & UI Rendering: Voila, Jupyter Widgets (
ipywidgets) for dynamic, real-time optimization visualization
├── GHG_Scope3_MultiTier_HGT.ipynb # Supply chain graph transformer architecture
├── EarthMAE_Self_Supervised_Final.ipynb # Geospatial self-supervised vision model
├── Dynamic_MACC.ipynb # Multi-period carbon abatement optimization
├── AI_LCA_fixed.ipynb # Multi-modal foundation model for generative LCA/PCF
├── requirements.txt # Project dependencies
└── README.md # Research Overview & Architectural Documentation