This report outlines the technical implementation of a multi-task Convolutional Neural Network (CNN) designed for 24-hour atmospheric forecasting using multi-year meteorological datasets (2018–2021).

The model, WeatherCNN, uses a deep spatial feature extractor coupled with a multi-task prediction head.

The backbone consists of sequential blocks designed to downsample spatial resolution while increasing feature depth:

• Input Stage: Accepts a 42-channel tensor representing various atmospheric variables (e.g., temperature, pressure, wind).
• Layer 1: ConvLayer (42 $\rightarrow$ 64 filters, $3 \times 3$ kernel) followed by BatchNorm2d and ReLU.
• Pooling: MaxPool2d ($2 \times 2$ kernel, stride 2) reduces the spatial dimensions.
• Layer 2: ConvLayer (64 $\rightarrow$ 128 filters, $3 \times 3$ kernel) followed by BatchNorm2d and ReLU.
• Global Pooling: A second MaxPool2d layer further abstracts the spatial features.

The model utilizes a LazyLinear layer in the FCLayer class. This allows the model to automatically infer the flattened input size from the convolutional backbone, outputting a 7-dimensional vector:

• Indices 0–5: Continuous variables (Regression).
• Index 6: Event probability (Classification).

Meteorological data often contains missing values that can destabilize training. This pipeline employs a three-tier cleaning strategy:

1. Scanning: A utility identifies "dirty" files containing NaNs across 42 channels.
2. Specific Logic: Channel 5 (DSWRF at surface) contains legitimate NaNs during nighttime. These are explicitly imputed with 0.0 within the WeatherDataset class.
3. Filtering: Any file containing NaNs in non-exempt channels is excluded from the training indices.

To ensure stable gradient descent, both inputs and targets are scaled:

• Inputs: Per-channel Z-score normalization using pre-computed means and standard deviations from the training set.
• Targets: Regression targets are normalized. During training, the model predicts in the normalized space, while "real-unit" MSE and RMSE are tracked for interpretability.

The model was trained on a single NVIDIA GPU using a multi-task loss approach.

The training objective minimizes a combined loss: $$L_{total} = L_{MSE}(\text{Regression}) + L_{BCEWithLogits}(\text{Classification})$$

Training logs below combine work from two setups: the original 42-channel report (Darius) and Subhanga-6 feature CNN training on the Tufts CS 137 cluster.

The model showed strong convergence within the first three epochs, particularly in the classification task.

Analysis:

• Rapid convergence: The model effectively learned the primary variance of the weather patterns within two epochs, cutting the real-unit RMSE by more than half.
• Stability: BatchNorm and AdamW prevented gradient spikes despite the high dynamic range of the input weather data.
• Best model: The lowest total loss ($0.5546$) was achieved at Epoch 2, after which the model was saved as the production checkpoint.

Checkpoint: checkpoints/best_model.pt.

• GPU: NVIDIA A100-PCIE-40GB, CUDA 12.9
• PyTorch (module): module load pytorch/2.8.0-cuda12.9-cudnn9
• Environment: conda activate base on the HPC; training driven from that stack instead of a one-off venv (avoids missing-package issues on compute nodes).

Training started ~9:13 AM targeting the A100 + CUDA 12.9 stack. Notable batch attempts:

1. Job 386265 (~9:27 AM) — virtual environment not installed on the node.
2. Node 386279 (~10:02 AM) — switched to conda after fixes; job 386281 (~10:08 AM).
3. Job 386303 (~10:37 AM).
4. torch.OutOfMemoryError on GPU (requested ~1.54 GiB; ~44 GiB already in use by PyTorch). Node 386627, job 386631 — led to batch/memory tuning (PyTorch docs suggest PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True when fragmentation is an issue).
5. Timeout on node 387630.
6. Job 387680 — further retry.

Final successful log excerpt below matches a completed run (e.g. subhanga/logs/train_388842.out); best weights written to checkpoints/subhanga/best.pt.

Python 3.11.13
torch 2.8.0+cu129
cuda available True
DATASET_DIR=/cluster/tufts/c26sp1cs0137/data/assignment2_data/dataset
CHECKPOINT_DIR=/cluster/tufts/c26sp1cs0137/supadh03/SkyOracle/checkpoints/subhanga
Device: cuda
Total time steps: 47952
Total usable inputs (t with t+24 in range): 47928
Train inputs: 21672  years=(2018, 2019, 2020)
Val inputs  : 8760  years=(2021,)
Targets shape: (47952, 6)
Binary label positives: 988/47952
Train inputs after NaN-target filter: 21144
Val inputs after NaN-target filter  : 8760
Loaded channel normalization stats from channel_stats.pt
Mixed precision (autocast): enabled, dtype=torch.bfloat16

Training completed with Done. Job finished: Thu Apr 2 02:58:42 PM EDT 2026.

Takeaways: Mixed precision (bfloat16 autocast), filtered NaN targets (21144 train / 8760 val), and channel stats from channel_stats.pt. Best validation MSE in this run was 0.718675 at epoch 5 (checkpoints/subhanga/best.pt).