Date: 5^th April, 2026

​

ANRF AISEHack

2026�Finale presentation | Theme : Pollution Forecasting

​

TeamName- Ctrl+alt+Defeat

TeamMembers-

Parthiv Kalikota

K V Sai Geetesh

Nihal Sareddy

Anjith Goud Koyyada

Shiva Sai Bangari

​

Your salient contributions

f1

Modelling strategy:

​

f2

Training:

f3

Results + Actions:

ANRF AISEHack 2026 Edition 1 – Phase 2 Finale, IIIT Hyderabad

• Physics aware ConvLSTM and UNet hybrid.
• Temporal modeling using wind and boundary layer height
• Spatial learning with Attention UNet

• Dynamic episode-aware loss
• Exponential moving average for stability
• Data augmentation with flips

​

• Strong validation performance
• Improved prediction of extreme pollution events
• Robust and scalable architecture

MATCHMAKING

Modelling - Strategies

01

Temporal Modeling

01

02

Spatial Modeling

• Attention-based UNet for spatial feature extraction
• Skip connections preserve fine spatial details
• Self-attention captures long-range dependencies

02

03

Physics-Aware Design

• Incorporates meteorological factors affecting pollution
• Focus on transport dynamics using wind fields
• Combines data-driven learning with physical intuition

03

Architecture diagram

​

Strengths:

• Captures both temporal and spatial dependencies
• Uses physically meaningful variables

Weaknesses:

• Requires careful tuning
• Computationally expensive

• Uses stacked ConvLSTM for sequence learning
• Captures evolution of pollution over time
• Inputs include PM2.5, wind components, and boundary layer height

REVIEW GROUNDING

Key Results – Why , what and how of the workings?

S1

What is working?

• Hybrid architecture combining ConvLSTM and Attention Unet
• Temporal modeling using wind and boundary layer height
• Episode-aware loss improving extreme event prediction
• Stable training using EMA and augmentation

● Implemented

All core setup criteria done

S2

How is it working? (Quantitative evals)

• Validation score: 0.924
• Strong episode correlation and lower SMAPE
• Consistent performance across timesteps
• Better spike prediction than baseline models

S3

Why is it working? (Reasoning and analysis)

• ConvLSTM captures temporal evolution
• Uses physically meaningful variables
• Attention improves spatial feature learning
• Focus on extreme events improves robustness

​

​

S4

• Hybrid ConvLSTM and Attention UNet architecture with high capacity.
• Trained on GPU with mixed precision for 80 epochs
• Uses EMA, augmentation, and dynamic loss for better convergence
• Shift from static CNN to physics-aware spatio-temporal modeling

​

MATCHMAKING

Key Results - Why , what and how of the workings?

Quantitative Performance & Benchmarks

Visualizations

MATCHMAKING

Details of Relevant Experimentation and Methodology

03 References�[1] OpenSTL Contributors. (2023). OpenSTL: A unified framework for spatio-temporal predictive learning [Software]. GitHub. https://github.com/chengtan9907/OpenSTL�[2] Woo, S., Park, J., Lee, J. Y., & Kweon, I. S. (2018). CBAM: Convolutional block attention module. In Proceedings of the European Conference on Computer Vision (ECCV).

01 Ablations and Key Findings�Ablation studies showed that temporal modeling using ConvLSTM significantly improved prediction of pollution spikes. Removing the episode-aware loss reduced performance on extreme events, while increasing model capacity enhanced spatial detail. These results highlight that temporal dynamics and event-focused training contributed most to overall performance.

02 External Resources and References�The approach is inspired by established architectures such as ConvLSTM for spatio-temporal modeling and UNet for spatial learning. Attention mechanisms like CBAM further improve feature representation, helping the model capture complex dependencies in pollution data.