🧪 Scientific-Mood ML Challenge

Scientific Modeling out of Distribution (Scientific-Mood)
NSF HDR ML Challenge — Modeling Out-of-Domain Extrapolation on Critical Scientific Processes

Machine learning models excel at interpolating across training datasets. This challenge asks models to extend beyond their training by performing out-of-domain extrapolation on practical, critical scientific processes that have not yet been well studied.

The HDR ML Challenge program presents three scientific benchmarks for modeling out of distribution in critical areas, plus one combined challenge.

📅 Challenge Period: September 18, 2025 — February 22, 2026 (11:59 pm AOE)

---

📂 Repository Structure

Scientific-Mood ML Challenge/
│
├── NSF HDR Scientific Modeling out of distribution NeuralForecasting/
│   ├── baselines/
│   │   ├── submissions/AMAG/        # Submission-ready model weights & code
│   │   └── training/AMAG/           # Training, evaluation & utility scripts
│   ├── CITATION.cff
│   ├── LICENSE
│   └── README.md
│
├── Beetles Scientific-Mood ML Challenge/
│   ├── submissions/                  # Inference & submission code
│   ├── training/                     # Training scripts & utilities
│   └── README.md
│
├── Predicting Coastal Flooding Events - Scientific Out-of-Distribution Challenge/
│   ├── submissions/xgboost_regression/   # Submission model & config
│   ├── training/xgboost_regression/      # Data conversion & training scripts
│   ├── LICENSE.txt
│   └── README.md
│
└── README.md                         # ← You are here

---

🧠 Sub-Challenge 1: Neural Forecasting

Forecast activations of a cluster of neurons from prior signals — vital for brain-chip interfaces and artificial limb control.

Item	Detail
Model	AMAG (Adaptive Multi-scale Attention Graph)
Key Techniques	RevIN, Multi-Scale Attention, Adaptive Graph Layer
Input Shape	(B, 20, N, F) — 10 observation + 10 prediction steps
Subjects	Affi (239 neurons) · Beignet (87 neurons)
Platform	Codabench — NeuralForecasting

Quick Start

# Install
pip install -r baselines/submissions/AMAG/requirements.txt

# Train
python baselines/training/AMAG/train.py --monkey affi --epochs 50
python baselines/training/AMAG/train.py --monkey beignet --epochs 50

# Evaluate
python baselines/training/AMAG/evaluation.py --monkey affi
python baselines/training/AMAG/evaluation.py --monkey beignet

📖 Full Details →

---

🪲 Sub-Challenge 2: Climate Prediction Using Ecological Data (Beetles)

Predict drought conditions (SPEI) over multiple timescales using images of ecological indicator organisms (ground beetles).

Item	Detail
Model	Hybrid Regressor (BioClip-2 ViT-B/16 + ConvNeXt Base)
Regression Head	Hidden Size: 512 → 256 → 6 with categorical metadata embeddings
Target	SPEI at 30-day, 1-year, and 2-year scales
Input Features	Beetle satellite images, scientific names, Domain IDs
Training	5-Fold CV · 50 epochs · AdamW + Cosine Annealing LR
Data & Weights	🤗 Hugging Face: jason79461385/beetles

Key Training Innovations:

• Extreme-Value Weighted Loss — (MSE × (1 + |target|)).mean() to emphasize extreme drought/wet conditions
• Layer Unfreezing (BioClip) — Only last 2 ResBlocks unfrozen to prevent catastrophic forgetting
• Dual Learning Rate (BioClip) — 10⁻⁵ for backbone / 10⁻⁴ for regression head

Quick Start

# Install
cd training
pip install -r requirements.txt

# Train (BioClip)
python train_bioclip.py

# Train (ConvNeXt)
python train_convnext.py

📖 Full Details →

---

🌊 Sub-Challenge 3: Coastal Flooding Prediction Over Time

Model sea levels at multiple sites across decades to predict coastal floods driven by climate change.

Item	Detail
Model	XGBoost Regression
Input Window	168 hours (7 days)
Metric	MCC (Matthews Correlation Coefficient)
Data Sources	Historical .mat station data (1950–2020, 12 stations)

Quick Start

# Install
pip install -r training/xgboost_regression/requirements.txt

# Generate training data
cd training/xgboost_regression/
python convert_to_parquet.py

# Train
python train_xgb_regression.py

# Evaluate
cd ../../submissions/xgboost_regression/
python model.py --test_hourly <path_to_test_hourly.csv> --test_index <path_to_test_index.csv> --predictions_out submission.csv

📖 Full Details →

---

🏆 Challenge Organizers

Imageomics

• Elizabeth G. Campolongo
• Wei-Lun Chao
• Chandra Earl (NEON)
• Hilmar Lapp
• Kayla Perry
• Sydne Record
• Eric Sokol (NEON)

Student Organizers: David E. Carlyn · Alyson East · Connor Kilrain · Fangxun Liu · Zheda Mai · S M Rayeed · Jiaman Wu

A3D3

• Yuan-Tang Chou
• Ekaterina Govorkova
• Philip Harris
• Shih-Chieh Hsu
• Mark S. Neubauer
• Amy Orsborn
• Leo Scholl
• Eli Shlizerman

Student Organizer: Jingyuan Li

iHARP

• Ratnaksha Lele
• Aneesh Subramanian
• Josephine Namayanja
• Bayu Tama
• Vandana Janeja

Student Organizers: Sai Vikas Amaraneni · Emam Hossain · Maloy Kumar Devnath · Subhankar Ghosh