Hila Paz Herszfang

While the world obsesses over LLMs, space missions still rely on decades-old linear compression methods like PCA. NASA’s PACE satellite scans Earth nearly 500,000 times per day, yet transmission bandwidth limits prevent its full hyperspectral richness from being downlinked.

In a Harvard–NASA collaboration originally focused on evaluating additional metaoptical hardware to enable PCA-based compression, we asked a different question: what if innovation happens in software instead? Rather than adding costly sensing gear, we designed an asymmetric autoencoder architecture tailored to existing satellite constraints. If successful, this approach reduces hardware dependency while enabling richer spectral retention.

This work reframes hyperspectral compression as a constraint-driven learning problem. We demonstrate that architectural choices determine which spectral signals survive compression, challenging reliance on predefined wavelength selections that were not optimized for learning.

The broader lesson is that meaningful AI impact does not always require larger models; sometimes it requires rethinking models to operate in alignment with physical constraints.