When a satellite or a laboratory sensor captures a pixel of data, it rarely records a "pure" substance. Instead, it captures a mixture—a 30-meter patch of Earth might contain grass, soil, and asphalt, all blended into a single spectral signature. The challenge of spectral unmixing is to reverse this process: to decompose a single signal into its constituent "endmembers" and their relative abundances.
While standard machine learning models are excellent at curve fitting, they often produce results that are mathematically valid but physically impossible. Recent research has shifted toward embedding physical laws directly into the optimization loop to ensure that "unmixed" data corresponds to reality.
The Geometry of the Simplex
The most fundamental physical requirement in unmixing is that the proportions of materials in a mixture must be non-negative and must sum to exactly 100%. In mathematical terms, this means the abundance vector must lie on a unit simplex.
Standard neural networks do not naturally respect these boundaries. As Hairong Qi and Ying Qu observe in their work on the Unsupervised Sparse Dirichlet-Net, unconstrained architectures can predict negative concentrations or "ghost" materials that don't exist in the scene. By using a Dirichlet distribution as a prior, they force the network to output abundances that are inherently non-negative and sum to one.
This "simplex constraint" acts as a powerful regularizer. Charis Lanaras and Konrad Schindler demonstrated that by enforcing these constraints within a coupled matrix factorization framework, they could fuse low-resolution hyperspectral data with high-resolution RGB images far more accurately than unconstrained methods. The constraints prevent the optimization from wandering into noisy, physically impossible regions of the search space.
Embedding Physics into the Optimization Loop
Beyond simple sums and signs, researchers are now embedding complex light-matter interactions into differentiable models. Purely statistical approaches often fail when environmental conditions change—a mineral looks different under Martian sunlight than it does in a Earth-bound lab.
To solve this "spectral variability" problem, John Janiczek and Suren Jayasuriya introduced differentiable programming for unmixing. Instead of just learning a mapping, their model incorporates an atomistic oscillator physics model. By making the physics itself differentiable, the system can "reason" about how light interacts with materials at a molecular level, allowing it to generalize to new environments (like the surface of Mars) where training data is non-existent.
Similarly, in the realm of art conservation, Su Yan and colleagues use physics-constrained matrix factorization to see beneath the surface of paintings. By fixing the intensity ratios between primary and secondary emission lines—a known physical constant of elements—their algorithm can resolve overlapping spectral peaks that would be indistinguishable to a standard signal processing algorithm.
Resolving the Conflict Between Sparsity and Physics
A recurring tension in this field is the conflict between sparsity (the idea that a pixel only contains a few materials) and the sum-to-one constraint. Mathematically, the penalties used to enforce sparsity (like L1 regularization) often interfere with the requirement that abundances sum to 100%.
Alexandre Zouaoui and Behnood Rasti addressed this by decoupling material selection from abundance estimation. Their non-convex optimization approach allows the model to pick a concise set of materials from a massive library without sacrificing the physical integrity of the final proportions.
In materials science, this concept is taken even further by incorporating the Gibbs Phase Rule. Stefano Ermon and Carla P Gomes developed the CombiFD framework, which treats thermodynamic laws as discrete combinatorial constraints. This ensures that the number of coexisting chemical phases in a sample remains within realistic limits, preventing the model from overfitting to noise by proposing an impossible number of underlying substances.
Domain Invariance through Shared Abundance
One of the most practical applications of these constraints is in transfer learning. If we train a classifier on one sensor, it usually fails on another because the raw spectral signatures change. However, the underlying physical proportions of materials should remain the same.
Ying Qu and Hairong Qi leverage this in their PCTL-SAS framework. By projecting raw data into a "shared abundance space" governed by non-negativity and sum-to-one constraints, they extract features that represent actual material fractions. Because these fractions are tied to physical reality rather than sensor-specific voltages, the model can generalize across different lighting conditions and sensors without needing new labeled data.