Foundation models struggle with extreme wildfire smoke prediction, study finds

A new study evaluated the generalizability of foundation models for predicting extreme PM2.5 concentrations from wildfire smoke, a critical public health challenge. Researchers compared six time series foundation model (TSFM) configurations against traditional models like LSTM and BiLSTM using a 12-year dataset from California. The findings indicate that while foundation models show modest improvements over naive persistence, they do not outperform fully trained recurrent baselines, and some exhibit instability in extreme conditions. Fine-tuning with LoRA significantly improved adapted foundation models but still fell short of the performance of trained recurrent networks. AI

IMPACT Challenges the assumption that larger foundation models universally outperform traditional methods in specialized environmental forecasting tasks.

RANK_REASON Research paper evaluating foundation models on a specific environmental forecasting task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• BiLSTM
• California
• Chronos-2 Forecasting Model
• electronic health records
• long short-term memory
• Moirai-2
• PM 2.5
• Time-MoE
• TimesFM
• Transformer++

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →