Date: 5th April, 2026
ANRF AISEHack
2026�Finale presentation | Theme : India in the Haze (PM2.5)
Raj Shah, Saptarshi Misra, Hiteshri Shastri | Team RuVision
Your salient contributions
f1
Architecture: ConvLSTM + 2-level UNet (0.6M params)
f2
Training: 28 epochs, AdamW, OneCycleLR, K=16 cluster, loss weights
f3
Score: 0.8636 | +0.0233 over previous v6
ANRF AISEHack 2026 Edition 1 – Phase 2 Finale, IIIT Hyderabad
ConvLSTM + UNet trained to forecast PM2.5 16h ahead from 10h lookback. Final score: 0.8636. Key innovations: simplified architecture (1-layer ConvLSTM, 2-level UNet, no CBAM), K=16 spatial clustering for loss weighting, and explicit episode-aware loss (α=1.0, β=3.0, γ=0.15).
MATCHMAKING
Modelling - Strategies
01
Architecture
1-layer ConvLSTM (dim=48) + 2-level UNet (base=24). No CBAM. 0.6M params.
01
02
Loss Function
α=1.0·GlobalSMAPE + β=3.0·EpSMAPE + γ=0.15·(1-EpCorr). Smooth SMAPE, pseudo-Huber δ=0.1.
02
03
Clustering
KMeans K=16 on PM2.5 + lat/lon. Weights from mean, episode freq, variance.
03
Input (10h, 20ch) → ConvLSTM → 2-level UNet → +Persistence → Forecast (16h)
Strengths: Lightweight (0.6M), fast training (~15 min/epoch on T4), no CBAM overfitting, episode-aware loss.
Weaknesses: Single model may miss regional variation. Regional ensemble attempted but did not improve.:
AdamW + OneCycleLR, 28 epochs, batch=16. 4 months WRF-Chem 2016.
REVIEW GROUNDING
Key Results – Why , what and how of the workings?
S1
What is working?
ConvLSTM+UNet (v7): 1-layer LSTM dim=48, 2-level UNet base=24, no CBAM. Loss: α=1.0·GlobalSMAPE + β=3.0·EpSMAPE + γ=0.15·(1-EpCorr). K=16 spatial clusters.
S2
How is it working? (Quantitative evals)
v6→v7 gains: Arch simplification +0.014, K=16 clustering +0.005, explicit β=3.0 episode loss +0.021. Combined score: 0.8636. Val metrics: ng=0.86+, ne=0.87+, nc=0.85+
S3
Why is it working? (Reasoning and analysis)
Simpler arch reduces overfitting on ~2000 training samples. Explicit EpSMAPE term directly optimises the hardest competition sub-metric. K=16 clusters give finer spatial loss weighting (IGP vs coast vs NE India).
S4
0.6M params | Training: ~15 min on T4 GPU, 28 epochs total | Inference: ~2s per test batch (TTA: E-W flip). Phase 1→2 shift: moved from IGP binary mask to K=16 data-driven spatial clusters; replaced episode ramp with explicit β·EpSMAPE term; reduced model capacity 5M→0.6M params to prevent overfitting.
MATCHMAKING
Key Results - Why , what and how of the workings?
01
Experiment Scores
v6 baseline: 0.8403 | Exp1-B (slim arch): 0.8542 | Exp3-B (K=16): 0.8449 | Exp4-D (β=3.0): 0.8608 | v7 combined: 0.8636 | Regional: 0.8552
01
02
v7 Architecture
1-layer ConvLSTM + 2-level UNet + no CBAM = 0.6M params vs v6’s 5M. Simpler model generalises better on ~2000 training samples.
02
03
Phase 1 → Phase 2
Phase 1 used a generic loss function with experiments on Huber, Focal and MSE. Phase 2 used score metric defined as loss function and fine tuned the weights.
03
Leaderboard score: 0.8636.
Score = w1·NormGSMAPE + w2·NormEpCorr + w3·NormEpSMAPE
Strengths: Fast convergence, lightweight, episode-aware loss directly aligns training with eval metric.
Weaknesses: Regional ensemble did not improve (0.8552). Could not experimented with lighter regional models due to time constraint.
Regional ensemble: 0.8552. Exp2/5/6 showed diminishing returns.
MATCHMAKING
Details of Relevant Experimentation and Methodology
01
Ablations (sequential, each vs v6=0.8403)
Arch -0.014, K=16 +0.005, β=3.0 +0.021. Exp2/5/6/regional: no gain.
01
02
References & Inspiration
ConvLSTM (Shi et al. 2015) for spatiotemporal forecasting. CBAM attention (Woo et al. 2018) — tested and found to overfit at this data scale. Pseudo-Huber smooth SMAPE for stable gradient near zero.
02
03
GenAI Tool Usage (Claude by Anthropic)
Claude used for: iterative notebook generation, loss function derivation from competition metric, and submission documentation. All model decisions validated empirically on leaderboard.
03
v6 → 4 ablation experiments → v7 → regional → presentation.