AI-in-Weather-and-Tokenomics-II--Exploring-Explainable-AI-for-Data-Valuation

Level: MA/MAP
Responsible person: Dr Mark C. Ballandies
Keywords:  AI, Machine learning, weather forecasting, Explainable AI, Data valuation, Tokenomics, DePIN

Artificial Intelligence (AI) continues to advance the skill and efficiency of weather prediction, with modern models matching or exceeding traditional numerical systems while using far less compute [1]. In parallel, decentralized physical infrastructure networks (DePIN) promise to crowdsource sensor data via token incentives, yet quantifying the value of contributed data (and defending against low-quality or adversarial inputs) remains a core challenge [2]. This project connects explainable AI for weather with principled data valuation to inform tokenomic reward design.   

You will build on our prior thesis that implemented a reproducible attribution pipeline and interactive dashboard for pretrained weather models (e.g., FCN, SFNO), and mapped where/which variables matter for near-surface temperature.  

Your focus (pick one, mix, or investigate own directions) 

• Faithfulness & robustness: Design causal/ablation checks that quantify how forecast skill degrades when “important” inputs are masked, shuffled, or replaced, establishing when an attribution is actually useful for this domain.  

• Global explanations: Move beyond single-pixel targets to regional/global objectives (e.g., area-averaged t2m, wind fields), producing global importance rankings across variables and locations.  

• Attribution-guided data valuation: Prototype source/region-level valuation using Shapley-style approximations or ablation deltas to score sensors/tiles by their marginal contribution to forecast skill, then translate scores into a rewarding rule. (See [3]).   

• Model & target breadth: Extend the pipeline to more variables (e.g., winds, humidity) and additional pretrained models where feasible, comparing stability across architectures.  

Why this matters 

A validated bridge from explanations → data value would enable transparent incentives for participatory sensing networks: reward the data that measurably improves forecasts, penalize what does not, and communicate why—all grounded in verifiable tests rather than intuition.  

What you’ll need: Strong ML background (PyTorch), and ideally comfort with gradient-based XAI (e.g., Integrated Gradients), or the drive to learn quickly.   

References: 
1]: Zhang, Y., Long, M., Chen, K., Xing, L., Jin, R., Jordan, M.I. and Wang, J., 2023. Skilful nowcasting of extreme precipitation with NowcastNet. Nature, 619(7970), pp.526-532.  
[2] Chiu, M.T., Mahajan, S., Ballandies, M.C. and Kalabić, U., 2025, June. Depin: A framework for token-incentivized participatory sensing. In 2025 IEEE International Conference on Blockchain and Cryptocurrency (ICBC) (pp. 1-7). IEEE. 
[3] Ghorbani, A. and Zou, J., 2019, May. Data shapley: Equitable valuation of data for machine learning. In International conference on machine learning (pp. 2242-2251). PMLR.  

References:

[1]: Ballandies, M.C., Carpentras, D. and Pournaras, E., 2024. DAOs of collective intelligence? Unraveling the complexity of blockchain governance in decentralized autonomous organizations. arXiv preprint arXiv:2409.01823. 
[2]: Helbing, D., 2012. Agent-based modeling. In Social self-organization: Agent-based simulations and experiments to study emergent social behaviour (pp. 25-70). Berlin, Heidelberg: Springer Berlin Heidelberg. 
[3]: Mariani, M.S., Ren, Z.M., Bascompte, J. and Tessone, C.J., 2019. Nestedness in complex networks: observation, emergence, and implications. Physics Reports, 813, pp.1-90.