Machine learning improves ocean drift forecasts by 50 percent when fed the right data

Researchers trained an AI system to predict how objects drift on the ocean surface, then tested which types of input data made it most accurate. Adding sea surface height measurements to current data cut prediction errors in half, while adding temperature data made things worse — suggesting that not all geophysical measurements help machine learning models equally.

Ocean drift prediction matters for search and rescue, pollution tracking, and shipping. The finding cuts through the assumption that feeding a machine learning system more data always improves results — it doesn't. This means oceanographers and maritime agencies may need to rethink which measurements they prioritize in their data pipelines, and it suggests that blindly adding sensors or variables to ML systems can actually degrade performance. The practical implication is straightforward: smarter input engineering beats more input.