Previous principal component analysis of ocean color absorption coefficient spectra 𝑎(𝜆) have shown the variation in these data is captured by a few eigenfunctions. Here, we perform an unsupervised, non-negative matrix factorization (NMF) of 𝑎(𝜆) to derive its fundamental and physically interpretable modes. When applied independently to two large datasets-one semi-empirical and one from inline measurements of the Tara Microbiome expedition-we find that four NMF basis functions describe > 99.9% of the variance in each. Furthermore, despite significant differences between the datasets in methodology and by geographic and temporal acquisition, their basis functions show very similar features at wavelengths 𝜆 ≈ 400 − 750 nm. Two of the modes capture the amplitude and spectral slope of absorption by color dissolved organic matter and/or detritus. The other two describe absorption by phytoplankton (𝑎 ph) separated into the pigments that couple tightly to the Chlorophyll-a 675 nm feature and another that captures 𝑎 ph variability at ≈ 450 nm. These two components provide an additive model for 𝑎 ph that outperforms the community-adopted non-linear model. Together, the majority of ocean color absorption is physically described by these four fundamental modes. Their coefficients follow generally understood biogeographic patterns for detritus and chlorophyll-a and highlight nuances in these patterns via the secondary modes. Conversely, spectra poorly modeled by these basis functions represent outliers that may reflect rare processes, instrument or processing issues, and unique pigment combinations. Another application is a simple, physically-interpretable representation of 𝑎(𝜆) for retrievals from hyperspectral, remote-sensing reflectances in the now active NASA PACE mission.  [Submitted to Limnology and Oceanography]