Differentiable ML Framework Exposes Hidden Trade-offs in LNP Characterization, Bridging SAXS Limits and mRNA Manufacturing Gaps

This arXiv preprint (v1, May 2026) introduces a differentiable machine-learning pipeline that accelerates small-angle X-ray scattering (SAXS) inversion for polydisperse lipid nanoparticles (LNPs) used in mRNA delivery. The authors combine a core-shell geometric model with Gaussian random-field interiors, replace the expensive monodisperse scattering calculation with a neural surrogate that speeds inference by four orders of magnitude, and add a differentiable integration layer over size distributions. Applied to both synthetic data and experimental MC3 LNPs, the method reveals that multiple distinct parameter sets can produce near-identical SAXS curves, with experimental fits dominated by a size-distribution versus interior-structure degeneracy. Unlike conventional non-differentiable fitting routines that explore only narrow parameter spaces, this framework enables large-scale multi-start optimization and ensemble identifiability analysis. The work remains a preprint and has not undergone peer review; the experimental component uses a modest number of MC3 LNP formulations without reported replicate counts or blinded validation sets, limiting statistical power. Mainstream coverage of AI in vaccines rarely connects these computational advances to the persistent manufacturing bottleneck of rapid, label-free LNP structural QC. Related studies on SAXS for LNPs (Yanez Arteta et al., 2018, Nanoscale) and differentiable programming for scattering (Liu et al., 2023, IUCrJ) underscore that traditional models remain too slow for real-time process analytics. The present framework directly addresses that gap by making high-throughput ensemble fitting tractable, yet the identifiability analysis also cautions that SAXS alone may never fully resolve core-shell versus disordered interior modes without orthogonal modalities such as cryo-EM or neutron contrast variation.