Machine learning-based inverse design for electrochemically controlled microscopic gradients of O2 and H2O2

Abstract

A fundamental understanding of extracellular microenvironments of O₂ and reactive oxygen species (ROS) such as H₂O₂, ubiquitous in microbiology, demands high-throughput methods of mimicking, controlling, and perturbing gradients of O₂ and H₂O₂ at microscopic scale with high spatiotemporal precision. However, there is a paucity of high-throughput strategies of microenvironment design, and it remains challenging to achieve O₂ and H₂O₂ heterogeneities with microbiologically desirable spatiotemporal resolutions. Here, we report the inverse design, based on machine learning (ML), of electrochemically generated microscopic O₂ and H₂O₂ profiles relevant for microbiology. Microwire arrays with suitably designed electrochemical catalysts enable the independent control of O₂ and H₂O₂ profiles with spatial resolution of ∼10¹ μm and temporal resolution of ∼10° s. Neural networks aided by data augmentation inversely design the experimental conditions needed for targeted O₂ and H₂O₂ microenvironments while being two orders of magnitude faster than experimental explorations. Interfacing ML-based inverse design with electrochemically controlled concentration heterogeneity creates a viable fast-response platform toward better understanding the extracellular space with desirable spatiotemporal control.

Keywords: O2 and H2O2 microenvironments; inverse design; microwire array; neural networks; spatiotemporal heterogeneity.

Fig. 1.

AI-based inverse design of electrochemically generated O₂ and H₂O₂ heterogeneities. (A) The ubiquitous spatiotemporal heterogeneities of O₂ and H₂O₂ in microbiology and the challenges posed in this research topic. (B) The combination of electrochemistry and ML-based inverse design offers a viable approach to mimicking and controlling the heterogeneities of O₂ and H₂O₂ in microbiology. O, oxidant; R, reductant; E_appl (t), the time-dependent electrochemical voltages applied on electrodes. (C) The design of the electrochemically active microwire array electrodes for the generation of O₂ and H₂O₂ gradients; 4e⁻ ORR & 2e⁻ ORR, four-electron and two-electron oxygen reduction reaction into H₂O and H₂O₂, respectively. (D and E) The 45°-tilting images of SEM for the representative microwire arrays used for the training of the ML model (D) and the ones inversely designed for targeted O₂ and H₂O₂ gradients (E); k = (P, D, L), the morphological vector that includes the P, D, and L of the synthesized wire arrays in units of micrometers. (Scale bars, 20 μm.)

Fig. 2.

Spatiotemporal control of O₂ gradient on Pt-loaded microwire array. (A) Design of Pt-loaded microwire array and the fundamentals of spatiotemporal mapping local O₂ concentrations ([O₂]) based on the intensity of phosphorescence emission (I_p) from Tris(1,10-phenanthroline)ruthenium(II) (Ru(phen)₃²⁺); ³O₂ and ¹O₂, the triplet and singlet dioxygen molecules, respectively; S₀ and S₁, the ground state and the first excited singlet state, respectively; T₁, the first excited triplet state; ISC, intersystem crossing; λ_ex and λ_em, the wavelengths of optical excitation and emission, respectively. (B and C) Cross-sectional I_p profiles on wire array k = (15, 4, 50) at t = 0, 16, and 48 s (B) and the subsequent temporal evolution of averaged local O₂ concentrations ([O₂]_avg) at different distances z from the base of the wire array (C). The values of E_appl are reported against RHE. The microwires are depicted in dashed lines in B. (Scale bars, 15 µm.) (D) Plots of [O₂]_avg versus z under different values of E_appl for wire array k = (30, 3, 50). (E) Plots of [O₂]_avg versus z in wire arrays of different k when E_appl = 0.5 V. Error bars represent SDs across multiple separate measurements in the device (n ≥ 3).

Fig. 3.

Spatiotemporal control of H₂O₂ gradient on Au-loaded microwire array. (A) Design of Au-loaded microwire array and the fundamentals of spatiotemporal mapping local H₂O₂ concentrations ([H₂O₂]) based on the intensity of fluorescence emission (I_f) in the fluorogenic reaction from Amplex Red to resorufin catalyzed by HRP. (B and C) Cross-sectional I_f profiles on wire array k = (15, 4, 50) at t = 0, 22, and 52 s (B) and the subsequent temporal evolution of the averaged local H₂O₂ concentrations ([H₂O₂]_avg) at different distances z from the base of the wire array (C). The values of E_appl are reported against RHE. The microwires are depicted in dashed lines in B. (Scale bars, 15 µm.) [H₂O₂]_avg = 0 when t = 0. (D) Plots of averaged local H₂O₂ concentrations ([H₂O₂]_avg) versus z under different values of E_appl for wire array k = (15, 4, 50). (E) Plots of [H₂O₂]_avg versus z in wire arrays of different k when E_appl = 0.5 V. Error bars represent SDs across multiple separate measurements in the device (n ≥ 3).

Fig. 4.

The development of inverse design for electrochemically generated O₂ and H₂O₂ gradients. (A) Comparison between the conventional protocol and our inverse design approach for the development of suitable experimental conditions, represented as {E_appl, k} in order to achieve desirable spatiotemporal distributions of O₂ and H₂O₂ concentrations ([O₂](r, t) and [H₂O₂](r, t), respectively). MLPNN 1 & 2, multiple-layer perceptron neural networks for O₂ and H₂O₂ gradients, respectively. (B) Protocols of data augmentation for the establishment of MLPNN; i_0,4e,Pt/Au and i_0,2e,Au, the exchange current densities of four-electron and two-electron ORRs on Pt and/or Au electrocatalysts, respectively. (C and D) The AMSE in the training (blue) and validation (pink) datasets at different epochs for the gradients of O₂ (MLPNN 1 in C) and H₂O₂ (MLPNN 2 in D). (E and F) Comparisons between the MLPNN-predicted values ([O₂]_predict and [H₂O₂]_predict) and training values ([O₂]_train and [H₂O₂]_train) for the local average concentrations of O₂ (E) and H₂O₂ (F), respectively. R², coefficient of determination.

Fig. 5.

Experimental validations of MLPNN-assisted inverse design. (A and B) Targeted O₂ and H₂O₂ gradients ([O₂]_avg(z) in A and [H₂O₂]_avg(z) in B, respectively) for exemplary inverse design assisted by the developed MLPNNs. (C and D) The three-dimensional contour plots of the probabilities that MLPNN-predicted [O₂]_avg(z) (C) and [H₂O₂]_avg(z) (D) match the targeted ones as a function of P, D, and L when E_appl = 0.5 V vs. RHE. (E and F) Experimental characterizations of [O₂]_avg(z) and [H₂O₂]_avg(z) (scattered points) in comparison with the targeted ones (lines) when {E_appl, k} = (0.5, 46, 6, 20) in E and (0.45, 17, 3, 30) in F on Pt- and Au-loaded microwire arrays, respectively. The potential applications of those yielded gradients in microbiology are noted. Error bars represent SDs across multiple separate measurements in the device (n ≥ 3).