Data-Driven Full-Spectrum Mapping of the Synthesis Landscape for Plasmonic Nanoparticles with High Chiroptical Activity

Technical Summary: Data-Driven Full-Spectrum Mapping of the Synthesis Landscape for Plasmonic Nanoparticles with High Chiroptical Activity

Problem Statement
Chiral plasmonic nanoparticles (NPs) offer significant potential for enantioselective catalysis, biosensing, and nanomedicine, yet achieving high chiroptical activity (quantified by the dimensionless asymmetry factor, or g-factor) remains a synthetic challenge. The synthesis of these NPs involves a high-dimensional, nonlinear chemical space where multiple parameters are coupled to morphology and optical properties. Conventional optimization approaches, which often rely on scalar targets like the maximum g-factor peak value, fail to reveal the distribution of different NP structures across the synthesis landscape or the mechanistic origins of chiroptical activity. Furthermore, samples with low g-factors are typically discarded during optimization, potentially obscuring intermediate growth states or competing pathways that are crucial for understanding the structure–property relationship.

Methodology
The authors established a data-driven framework to map the synthesis landscape of glutathione (GSH)-mediated gold 432 helicoid III nanoparticles. The approach integrated three core components:

1. Autonomous Synthesis: A custom-built automated liquid-handling platform, controlled by the Gryffin Bayesian optimization algorithm, executed 80 independent synthesis experiments within 30 hours. The system varied four reagent concentrations: Gold(III) chloride trihydrate (HAuCl₄), cetyltrimethylammonium bromide (CTAB), L-ascorbic acid (AA), and GSH.
2. Full-Spectrum Data Acquisition: Instead of recording single peak values, the team acquired full g-factor spectra (400–1100 nm, 351 data points) for all 80 samples.
3. Manifold Learning and Clustering: The high-dimensional spectral data underwent a three-step unsupervised workflow:
   • Functional Smoothing: Spectra were transformed into continuous functional objects using cubic B-splines.
   • Dimensionality Reduction: Functional Principal Component Analysis (fPCA) compressed the data into three principal scores, capturing ~98% of the variance.
   • Topological Mapping: K-means clustering (k=4) and Uniform Manifold Approximation and Projection (UMAP) were used to embed the data into a 2D topological landscape, revealing distinct structural branches.
4. Interpretable Machine Learning: The authors employed Gaussian process regression, SHapley Additive exPlanations (SHAP), logistic regression, and symbolic regression (using the SISSO algorithm) to correlate synthesis parameters and geometric features with the observed spectral outcomes.

Key Results

• High-Performance Synthesis: The autonomous optimization successfully identified reaction conditions yielding a maximum |g-factor| of 0.24, representing a 2.2-fold increase over previously reported empirical optima.
• Four-Branch Landscape Topology: The UMAP embedding revealed a non-continuous, four-branch topology rather than a graded landscape.
  • High-|g-factor| Clusters (C1–C3): These clusters correspond to 432 helicoids with varying degrees of structural complexity (from shallow twisted grooves to deep kinked crevices and fractures).
  • Low-|g-factor| Cluster (C0): This cluster exhibited significant heterogeneity, containing previously unreported chiral rhombic dodecahedra, cubes with shallow indentations, and other intermediate morphologies.
• Parameter Control Logic:
  • HAuCl₄ Loading: Identified as the dominant global variable determining access to the high-|g-factor| region. Higher concentrations (0.25–0.40 mM) were required to enter the high-performance branches.
  • GSH Loading: While having a weak global correlation with the scalar g-factor, GSH concentration was the primary discriminator within the high-|g-factor| region, conditionally selecting between the C1, C2, and C3 structural branches.
• Geometric Descriptor: Symbolic regression identified an empirical descriptor, d⁴L (where d is gap depth and L is particle edge length), which correlates linearly with the |g-factor| (|g-factor| = 1.831 × 10⁻¹¹ × d⁴L + 0.056). This highlights that chiroptical activity is highly sensitive to gap depth (fourth-power dependence) rather than just particle size.
• Discovery of Intermediate Structures: Analysis of the low-signal C0 region revealed chiral rhombic dodecahedra that bridge achiral octahedral seeds and the final 432 helicoids. The study proposes a growth model where these structures evolve through facet transitions ({111} → {110} → {100}). Notably, a subgroup (C0-a) exhibited chiroptical sign inversion, attributed to the dominance of threefold-symmetric features over fourfold helicoid features at low HAuCl₄ concentrations.

Significance and Claims
The paper claims that utilizing full g-factor spectra as coordinates, rather than scalar peak values, allows for the reconstruction of a complete parameter–structure–property map. This approach shifts the role of machine learning in nanomaterial synthesis from merely locating isolated performance optima to generating interpretable chemical-space maps.

The authors emphasize that retaining low-signal samples is critical; discarding them would have hidden the chiral rhombic dodecahedra, which are essential for understanding the mechanistic pathway from achiral seeds to high-performance helicoids. The study demonstrates that synthesis parameters do not act as independent scalar controls but regulate the g-factor by enabling access to specific structural branches or selecting among them. This methodology is presented as a generalizable strategy for other nanomaterial systems where optical responses encode morphology-sensitive information, facilitating a transition from optimization to mechanistic understanding.