Introduction

In the past 30 years, transition metal-catalyzed reactions have been widely used in organic synthesis [1]. In particular, cross-coupling, oxidation, addition, and metathesis reactions have been reported with excellent yields, selectivities, and functional group tolerance; these parameters have been used for the preparation of compound libraries and large-scale pharmaceutical production [2]. Suzuki–Miyaura coupling reactions are powerful tools for the formation of carbon‒carbon bonds because of their mild reaction conditions and easy separation of boron-containing byproducts [3]. Phosphines are representative ligands for Suzuki–Miyaura coupling reactions, and their reactivities and selectivities can be adjusted by tuning their steric and/or electronic properties [4].

In recent years, there has been growing interest in solid-supported catalysts because of their insolubility in the reaction solution, which facilitate easy separation and recycling from a green chemistry perspective [5]. Generally, the catalytic efficiencies of the supported catalysts are lower than those of their homogeneous counterparts, mainly due to the limitations of the mass transfer in the solid phase of the supported catalysts. To improve mass transfer in the supported catalysts, the fabrication of a porous morphology on the support material is beneficial. Ma et al. achieved excellent reactivity and alkene selectivity in the semi-hydrogenation of terminal alkynes using a Pd-immobilized porous polymer catalyst [6]. Due to their easy preparation and high chemical stability, transition metal catalysts supported on porous polystyrene have been developed [7]. Matsumoto et al. developed porous polystyrene monoliths supporting phosphine–Pd complexes via Suzuki–Miyaura coupling and reported that the large surface area and high ratio of monoligated phosphine–Pd complexes improved the catalytic efficiency. In particular, the degree of cross-linking of polystyrene significantly influenced the coordination behavior of the phosphine–Pd complexes [8]. To achieve the highest activity for these polymer catalysts, a significant number of trial-and-error experiments are needed. This process consumes a substantial amount of time and experimental resources. Optimizing catalysts manually through human effort is challenging due to the extensive experimental workload and associated costs. Therefore, automated learning and optimization based on existing data can enable substantial cost savings.

Machine learning is an information science that elucidates trends based on data and predicts properties [9]. With the remarkable development of machine learning methodologies in recent years, their applications in materials science have increased [10]. The application of machine learning in the field of polymers has gradually gained attention and made significant strides over the past five years, primarily due to the inherent complexity of polymers, the lack of appropriate descriptors, and the absence of extensive publicly available databases [11]. However, the use of machine learning for the optimization of immobilized catalysts on synthetic polymers has not yet been reported. Due to the complexity of the material’s preparation, collecting a large amount of data manually for building a machine-learning model is challenging.

Herein, we first report a computer-guided and systematic approach to optimize polymer-supported catalysts in transition metal catalysis (Fig. 1). As a machine learning method, we adopted Bayesian optimization, which leverages existing predictive models to suggest the next sampling point and enables faster convergence to the desired maximum value. The Bayesian optimization sample points are based on proposals provided by an acquisition function; this function balances exploration and exploitation to efficiently find the global maximum [12]. Applying Gaussian processes as surrogate models in Bayesian optimization is advantageous for quantifying uncertainty in the exploration process and can achieve predictions of reaction yields. Consequently, Bayesian optimization achieves optimization goals with a minimal amount of experimental data [13]. Moreover, researchers have applied Bayesian optimization to catalytic reactions [14, 15] and catalyst materials [16, 17].

Fig. 1
figure 1

Workflow of the machine learning-guided optimization of the Pd-immobilized porous polymer catalysts. The monolith polymerization step involved two independent variables of the DVB content (x) and 1-decanol content (y) to maximize the TOF as a target variable in the Suzuki–Miyaura coupling reaction. Bayesian optimization was applied for predictive modeling, and the optimized conditions were experimentally validated in subsequent iterations

Full size image

In this report, porous polystyrene monoliths supporting phosphine ligands with different cross-linking densities and porous structures were synthesized under various polymerization conditions. The monoliths were obtained as self-standing materials during polymerization-induced phase separation in a porogenic solvent (toluene/1-decanol) [8]. The cross-linker and porogen contents were selected as independent variables that influence the morphology and properties of the pores. Additionally, the cross-linking monomer also impacts the coordination state. Therefore, selecting the divinylbenzene (DVB) and 1-decanol contents as the variables were appropriate. The turnover frequency (TOF) was selected as the target variable for modeling via machine learning. Through Bayesian optimization of the dataset using a Gaussian process, the points with predicted maximum TOFs were proposed and experimentally evaluated. This experimental validation and iterative machine learning cycles contributed to the gradual improvement in the catalytic activity. This methodology combined experimentation and intelligent data-driven optimization for enhanced catalyst performance.

Experimental

Materials

Water with a conductivity of 18.2 MΩ cm (Milli-Q, Millipore Co., Bedford, MA, USA) was used in all experiments. Styrene (St, >99% GC purity, stabilized with 4-t-butylcatechol, Tokyo Chemical Industry Co., Ltd., Tokyo, Japan) and DVB (m- and p-mixture, containing ethylvinylbenzene and diethylbenzene, stabilized with TBC, Tokyo Chemical Industry Co., Ltd., Tokyo, Japan) were purified using a basic alumina column [19]. 2,2′-Azobis(isobutyronitrile) (AIBN, FUJIFILM Wako Pure Chemical Corporation), 4-(diphenylphosphino) styrene (VPP3, 97%, Sigma-Aldrich), bis(benzonitrile)palladium(II) dichloride (Pd(II), >98.0%, Tokyo Chemical Industry Co., Ltd., Tokyo, Japan), 1-decanol (99%, Tokyo Chemical Industry Co., Ltd., Tokyo, Japan), toluene (99.5%, FUJIFILM Wako Pure Chemical Corporation), tetrabutylammonium hydroxide (TBAOH, 10% in methanol, Tokyo Chemical Industry Co., Ltd., Tokyo, Japan), phenylboronic acid (containing varying amounts of anhydride, Tokyo Chemical Industry Co., Ltd., Tokyo, Japan), 1-iodo-4-nitrobenzene (>99%, Tokyo Chemical Industry Co., Ltd., Tokyo, Japan), 1-bromo-4-nitrobenzene (>99%, Tokyo Chemical Industry Co., Ltd., Tokyo, Japan), and nitrobenzene (NO2Ph, min 99.5%, Wako Pure Chemical Industries, Ltd., Osaka, Japan), 4-chloronitrobenzene (>98%, Tokyo Chemical Industry Co., Ltd., Tokyo, Japan) and tetrahydrofuran (THF, Super Dehydrated, with Stabilizer, Wako Pure Chemical Industries, Ltd., Osaka, Japan) were used.

Synthesis of Monoliths [8]

Monolithic porous polystyrene was prepared by the bulk copolymerization of VPP3 (2 wt%), DVB (x wt%) and St (100 − x − 2 wt%) in a porogenic solvent. The monomers were dissolved in a mixture of 1-decanol (100 − y wt%) and toluene (y wt%), followed by the addition of AIBN (1 wt% with respect to monomers) to produce a monomer solution (1 g, 40 wt%) (Scheme 1). The monomer solution was prepared in a glass vial. After degassing via three freeze‒pump–thaw cycles, the vial was charged with N2 and subsequently sealed with a screw cap. After the vial was carefully crushed, the resulting polymer was cut into ~5 mm small blocks and washed by immersion in anhydrous THF for more than 1 day, and the THF was changed three times. The monoliths were dried in vacuo at room temperature overnight. The internal structures of the dried monoliths were observed using FE-SEM.

Scheme 1
scheme 1

Synthesis of a porous polymer with a phosphine liganda. aConditions: VPP3 (2  wt% of monomer), DVB (x  wt% of monomer), St (100 − x − 2  wt% of monomer), 1-decanol (y  wt% of solvent), toluene content (100 − y  wt% of solvent), AIBN (1 wt% of monomer). Total monomer conc.: 40 wt%, 70 °C, 48 h

Full size image

Pd coordination

In a glove box, 2.3 mg of [PdCl2(PhCN)2] (0.006 mmol) was dissolved in anhydrous THF (6 mL). Subsequently, 1 mL of the Pd solution (0.001 mmol Pd) was added to the dried and cubic monolith (144 mg, 0.01 mmol of P, P/Pd 10:1) in a glass vial. The mixture was incubated at room temperature for more than 4 h, and the THF-swollen monolith turned yellow. After removal from the glove box, the Pd-loaded monolith was washed with anhydrous THF (2 mL). The absorbance of the reaction mixture was measured using ultraviolet‒visible (UV‒vis) at 440 nm to confirm the complete coordination of Pd(II) (Scheme 2).

Scheme 2
scheme 2

Ligand exchange reaction of Pd(II)

Full size image

Suzuki–Miyaura coupling reaction using Pd-loaded monoliths

The Suzuki–Miyaura coupling reaction between p-halogenated nitrobenzenes and phenylboronic acid was conducted in the presence of TBAOH in a THF/MeOH mixture. Initially, a THF reaction mixture was prepared in a glove box: p-halogenated nitrobenzenes (0.25 mmol), phenylboronic acid (45.5 mg, 0.375 mmol), and NO2Ph (20 μL, used as an internal substance) were dissolved in anhydrous THF (5 mL). Subsequently, a portion of the substrate solution (3 mL, containing 0.15 mmol 1-bromo-4-nitrobenzene at 1 eq; 0.225 mmol phenylboronic acid at 1.5 eq; 12 μL of NO2Ph) was added to a glass vial containing a Pd-loaded monolith (0.001 mmol Pd, 0.67 mol%, P/Pd = 10:1). Next, TBAOH (10% in MeOH, 4.5 mL, 0.45 mmol, 3 eq) was added to the reaction mixture, and the glass vial was purged with nitrogen. The reaction was initiated in a water bath at 40 °C, and a small aliquot was collected using a syringe for various reaction times and then analyzed using HPLC to determine the yield of 4-nitrobiphenyl (Scheme 3).

Scheme 3
scheme 3

Pd-catalyzed Suzuki–Miyaura coupling reaction of p-halogenated nitrobenzenes

Full size image

Results and discussion

Preparation and characterization of porous polymer-immobilized catalysts

Our research began with the synthesis of the porous polymer monoliths containing phosphine ligands. The radical polymerization of St, DVB, and VPP3 in a mixed solution of toluene and 1-decanol produced a monolithic and porous structure due to polymerization-induced phase separation [8]. The obtained monoliths were cut into pieces, washed by immersion in anhydrous THF, and dried under vacuum. SEM images confirmed that the monolith exhibited a porous structure, with pore sizes ranging between those of macro- and mesopores (5–600 nm, Fig. 2a and Fig. S1). Through mercury intrusion experiments, the pore sizes and distribution of each monolith were determined, and the pore sizes ranged from 5 to 263 nm (Fig. 2b, Fig. S5 and Table S1). The heterogeneous pore size distribution was likely related to the exothermic nature of the polymerization process [18]. Catalyst entry 7 exhibited multiple pore size distributions; thus, no evidence was observed to indicate that the heterogeneous pore size distribution adversely affected the catalyst activity. The results from solid-state NMR measurements confirmed that the polymerized monolith contained free P (Fig. 2c and Fig. S6). In the NMR spectrum, two peaks were observed at chemical shifts of 23 ppm and −6 ppm, indicating the presence of oxidized P and free P, respectively [8]. The dried monolith was immersed in a THF solution of [PdCl2(PhCN)2] (P/Pd = 10:1) for complexation between the polymer-supported phosphine and Pd. After 4 h, the absence of free Pd in the solution was confirmed using UV‒vis spectroscopy; thus, the immobilization of Pd on the monoliths was successful. The solid-state 31P NMR spectrum of the monoliths after Pd loading showed a peak in the vicinity of 24 ppm, accompanied by a decrease in the peak area of free P (Fig. 2c and Fig. S6). Therefore, the Pd coordination exclusively yielded P–Pd bis-complexes (24 ppm) rather than mono-coordinated Pd [8]. These results demonstrated the successful immobilization of Pd–P complexes onto polystyrene with varying polymer compositions and porous structures.

Fig. 2
figure 2

Properties of polymer 7. a SEM image, b pore size distributions, and c 31P CP/MAS NMR spectra before and after the catalytic reaction

Full size image

Catalytic activity for the Suzuki–Miyaura coupling reaction and Bayesian optimization

The catalytic activities of the Pd-immobilized monoliths were evaluated in Suzuki–Miyaura coupling reactions. The initial dataset included five different compositions of monoliths, where the DVB and 1-decanol contents were varied (polymer 1–5, entries 1–5, Table 1). The Suzuki–Miyaura coupling reaction between 1-iodo-4-nitrobenzene and phenylboronic acid was conducted at 40 °C. Small aliquots of the reaction mixture were collected at 5 min and analyzed using high-performance liquid chromatography (HPLC) to determine the TOF of the catalyst as a target variable (entries 1–5, Table 1). From the time-yield profile of the catalytic reactions, all reactions were completed within 20 min (entries 1–5, Fig. S4).

Table 1 The activity of the Suzuki–Miyaura coupling reactions catalyzed by Pd-immobilized porous polymer catalystsa
Full size table

Machine learning was employed using the GPy [20] model in Python to conduct the Bayesian optimization of the polymerization conditions (DVB content (x) and 1-decanol content (y)) for higher catalytic activities. Gaussian kernel functions and Gaussian processes were utilized during the optimization process. Bayesian optimization was performed in four steps: (i) the Gaussian process was used as a surrogate model, and the model was established with the initial data; (ii) the predicted average and confidence were acquired from the model, and the next experimental point was provided based on the maximized acquisition function, which was the upper confidence bound (UCB) in the present study; (iii) the experiment was performed with the next selected sampling point; and (iv) new experimental points were added, and optimization continually updated the model. The initial Bayesian optimization was conducted with 5 initial data points (entries 1–5, Table 1), and a contour map of the model is shown in Fig. 3a. The color gradient represents the TOFs, with red indicating higher values and blue indicating lower values. This visualization effectively represents the machine learning model utilizing Bayesian optimization of two variables. Subsequently, the contour map was sliced along both the horizontal and vertical axes at the predicted points corresponding to higher TOFs (red regions), resulting in two different sectional views (Fig. 3b–e). Due to the separation of the two red regions in the initial contour map, we generated four sectional views corresponding to each red region. In the sectional view, the dark blue line and light blue area indicate a predicted mean and a 25–75% prediction confidence interval. To determine the predicted maximum TOFs, we selected the upper limit of the confidence interval as the maximum point. This method of sampling points based on proposals can be described using the UCB acquisition function. The maximum UCB acquisition function was determined based on observing the upper limit of the 50% confidence interval. The maxima in the upper red region corresponded to the DVB and 1-decanol contents of 61 and 66 wt%, respectively (Fig. 3b, c). Similarly, in the lower red region, the optimized DVB and 1-decanol contents were 39 and 24 wt%, respectively (Fig. 3c, d).

Fig. 3
figure 3

1st Gaussian process regression GPy for entries 1–5 in Table 1. a Estimated yield and b predicted yield for the DVB content in the red ring above in (a). c Predicted yield for the 1-decanol content in the red ring above in (a). d Predicted yield for the DVB content in the red ring below in (a). e Predicted yield for the 1-decanol content in the red ring below in (a)

Full size image

According to the Bayesian optimization, the monoliths were prepared under the optimal polymerization conditions (polymers 6 and 7, entries 6 and 7, Table 1), and their TOFs were evaluated in the Pd-catalyzed cross-coupling reactions (entries 6 and 7, Table 1). Notably, a significant improvement in the catalytic activity was observed for polymers 6 and 7, which produced TOFs of 20.7 and 24.3 min−1, respectively (entries 6 and 7), compared with those of the initial dataset (TOFs of 12.3–17.7 min−1, entries 1–5, Table 1).

These data (entries 1–7) were added to the dataset and subjected to 2nd-round machine learning using Bayesian optimization to obtain the polymer compositions (polymers 8 and 9, entries 8 and 9, Table 1) that could generate higher activity (Fig. S2). However, these catalysts yielded smaller TOF values (3.3 and 5.7 min−1, entries 8 and 9, Table 1), and longer reaction times were required to complete the catalytic reaction (entries 8 and 9, Fig. S3). Since no improvement was obtained in the second optimization, we determined that entry 7, which had the highest TOF (TOF = 24.3 min−1, entry 7, Table 1), was the optimal catalyst based on the research goal of reducing the number of experiments. This outcome demonstrated the success of Bayesian optimization in polymeric immobilized catalysts and established a direct correlation between the variables involved in the polymer preparation and the desired final properties while disregarding intermediate complex variables. This simplification of the optimization process led to the successful development of polymer-supported catalysts with minimal time and experimental costs.

Bayesian optimization has been applied in the polymer field, especially for optimizing polymerization processes and enhancing polymer performance. In processes for polymeric materials, the shape of the fibers [21], producibility [22], and polymerization kinetics [23] has been optimized by Bayesian optimization. On the other hand, designing and improving the sophisticated functions of polymers are tedious and time-consuming. In this sense, several functional polymers have been improved using Bayesian optimization for electrical insulating performance [24], water permeability and selectivity [25], mechanical/thermal properties [26], antimicrobial efficiency [27], and abrasion tolerance [28]. In this research, we investigated porous polymer monoliths with diverse structures and morphologies for high catalytic performance. Thus far, research applying Bayesian optimization to polymer catalysts is lacking. This study represents the first endeavor to apply Bayesian optimization to the polymer composition, successfully optimizing porous heterogeneous polymer materials.

Characterization of the Pd nanoparticle formation during the catalytic reaction progression

Before and after the catalytic reaction, the monolith was analyzed using scanning transmission electron microscopy (STEM) and energy-dispersive X-ray spectroscopy (EDX). Before catalytic reactions, Pd was inferred to be atomically dispersed on the monoliths without forming nanoparticles (Fig. 4a). After the catalytic reaction, Pd nanoparticles formed (Fig. 2a, b). The size distributions of the Pd were similar (4–15 nm) among the best and worst monolith catalysts (entries 7 and 8, Table 1).

Fig. 4
figure 4

Properties of polymer 7. a Bright-field STEM image after Pd coordination. b Catalytic reaction. c EDX mapping for Pd

Full size image

The relationship between the porous properties and reaction activities was evaluated and analyzed (Table S3); however, no correlation was observed for these properties and their reaction activities (see the ESI for details). Since all Pd species on the monoliths were bis-P-ligated, the coordination state was not a factor for the variations in the catalytic activity. Based on the STEM images (Fig. 4b, c and Figs. S7, S8), minimal differences were observed in the particle size of Pd among the monoliths. The formation and growth of nanoparticles were unrelated to their catalytic activities. Therefore, a hypothesis was proposed as follows: the process by which Pd aggregates formed particles was correlated with the extent of the catalytic reaction. Since both the best and worst monoliths showed the same TON, the resulting particle sizes were roughly similar. Thus, the aggregation of Pd particles was a consequence of the catalytic reaction rather than the cause of the difference in the reaction activity.

Performance of this catalyst with various reactants

We applied the best catalyst to Suzuki–Miyaura coupling reactions of different p-halogenated nitrobenzenes (X = Br or Cl). The reaction kinetic curves of the catalytic reactions revealed that the yields of 4-nitrobiphenyl (78% at 140 min) and TOF (1.5 min−1) were moderate when using 1-bromo-4-nitrobenzene (entry 10, Table 1). On the other hand, 4-chloronitrobenzene had a lower reactivity (3% at 150 min) (entry 11, Table 1). The reactivity of the aromatic halides followed the trend I > Br > Cl, as reported in the literature. As reported in the literature, aromatic chlorides do not undergo Suzuki–Miyaura coupling reactions, and Suzuki–Miyaura coupling reactions involving aromatic bromides require higher reaction temperatures and longer reaction times than aromatic iodides [29]. However, the catalyst employed in this study facilitated Suzuki–Miyaura coupling reactions for both aromatic iodides and aromatic bromides. Moreover, the reaction conditions were relatively mild, and the reaction proceeded at a fast rate.

Conclusions

In summary, the machine learning method of Bayesian optimization was utilized to optimize the composition of the polymeric porous immobilized catalysts and the composition of the porogenic agents. Following the optimization, the catalyst exhibited enhanced reaction activity (the TOF increased from 17.7 to 24.3 min−1). In this study, an approach to optimize polymeric catalysts using machine learning was presented, and a direct connection between the dependent and independent variables without considering the influence of intermediate variables was established. Moreover, the application of Gaussian process-based Bayesian optimization provided the advantage of optimization with minimal data and significantly simplified the originally complex optimization process in catalyst chemistry and materials science.