Extended Data Fig. 5: Flowchart of the holographic trimming algorithm.

Trimming holograms are formed with weighted Gerchberg-Saxton (GS) algorithms and projected onto desired cavities for duration Δt with power \({P}_{{{{\rm{trim}}}}}\). Alternating trimming and resonance readout periods continue until the instantaneous wavelength λ_i of any targeted cavity blueshifts past the target wavelength λ_t. Thereafter, a new set of target cavities is selected and trimmed. This selection and trimming sub-loop continues until all resonant wavelengths {λ₀} are below the ‘rest’ wavelength λ_rest, at which point trimming is halted and the resonances are continuously monitored at readout interval Δt_rest. When the resonances are sufficiently stable (redshifting from moisture adsorption to the silicon membrane is arrested), the total ‘rehydration’ redshift Δλ₀ of each cavity is updated to better estimate the true resonant wavelength λ₀ ≈ λ_i + Δλ₀ from the instantaneous wavelengths {λ_i} during trimming. The entire process terminates when the peak-to-peak static resonant wavelength uniformity \(\Delta {\lambda }_{0}^{{{{\rm{p-p}}}}}\) drops below the desired tolerance Δλ_tol.