Combining High‐Throughput Experimentation and Automated Flow for Bayesian Optimization of a Metallaphotoredox Reaction

ABSTRACT

In this work, a metallaphotoredox C(sp²)–C(sp³) coupling between aryl (pseudo)halides and alkyl boronic acid pinacol esters was optimized using a two‐stage workflow. In the first stage, high‐throughput experimentation (HTE) is used in conjunction with Bayesian optimization (BO) to select promising discrete parameter values for five categorical reaction variables, including photocatalyst, nickel precursor, nickel ligand, solvent, and amine, from a search space of 30,000 conditions within 120 experiments per substrate across four aryl (pseudo)halides, representing <0.5% of the space. Crucially, the approach proved able to autonomously identify reactivity trends reported in literature without any prior knowledge. In the second stage, the optimal conditions for each substrate are transferred to flow, where eight continuous parameters, including residence time, reaction stoichiometry, temperature, and light intensity, are fine‐tuned. The sequential strategy delivers DMF‐ and precious‐metal‐free protocols for bromides and chlorides and, for the first time, extended this coupling reaction to aryl triflates.

A two‐stage workflow combines plate‐based HTE optimization of five discrete reaction parameters with fine‐tuning of eight continuous variables in continuous flow to accelerate development of a metallaphotoredox C(sp²)─C(sp³) coupling. This hybrid approach identifies substrate‐specific optima, enables robust DMF‐ and precious‐metal‐free protocols and, for the first time, expands the reaction to aryl triflates.

1. Introduction

The pharmaceutical, fragrance, and fine‐chemical industries are increasingly confronted with significant challenges as they seek to modernize and adapt to a rapidly changing landscape. The pursuit of greener, more resource‐efficient synthetic routes has converged with regulatory and societal pressure to phase out hazardous solvents and reagents, and diminish reliance on precious transition‐metal catalysts, as evidenced by the European ‘Registration, Evaluation, Authorisation, and Restriction of Chemicals’ (REACH) regulation [1, 2].

Additionally, the industry is under increased pressure to shorten discovery and development times, while simultaneously tackling the complexities associated with designing more intricate molecules [3]. Together, these drivers have created fertile ground for automation‐enabled, data‐efficient optimization strategies, giving companies the means to balance innovation and efficiency. In this context, automated and data‐driven methodologies that can rapidly traverse high‐dimensional reaction space—without imposing prohibitive experimental burden—offer a compelling route to accelerate process development.

High‐throughput experimentation (HTE) is one of those methodologies, allowing chemists to rapidly investigate a large range of parameters while reducing material consumption. After the adoption by the large chemical companies, and thanks to the deployment of new and much more accessible tools, it has also been embraced by small‐to mid‐sized companies as well as academia and has become a key technology for the chemical industry in general [4]. HTE's main strength is parallelization through the use of multi‐well reaction plates (e.g., 96 or 384 wells), which greatly benefits manipulation of discrete variables, such as solvents, catalysts, or ligands [5]. However, this strength becomes a limitation when optimization focuses on continuous variables, like temperature or pressure, since it is not possible to independently vary these across individual wells on a plate using standard equipment [6].

Automated flow reactors can overcome this constraint and have attracted considerable attention for accelerating reaction optimization. Due to their ease of automation, continuous‐flow reactors excel where plate‐based HTE struggles. Variables like temperature, residence time, or reagent equivalents can be easily varied per individual experiment, making them ideal for fine‐grained exploration of continuous parameters [7]. Flow chemistry also offers engineering benefits—superior mass and heat transfer, precise residence time control, and improved light penetration—further enhancing its utility for complex transformations, like photochemical and photoredox reactions [8, 9].

In parallel with these hardware advancements, algorithmic reaction optimization has matured into a powerful approach for navigating complex, multidimensional design spaces. Several classes of algorithms have been used for this, but Bayesian optimization (BO) algorithms now dominate the field [10, 11]. BO leverages a probabilistic surrogate model of the objective function, combined with an acquisition function, to select the next experiment, balancing further exploration of the chemical space with exploitation of the best known conditions [12]. In simpler terms, BO uses the outcome of previous experiments to build a predictive model of the reaction space, which is then used to decide which experiment is most informative to perform next. The result is a data‐efficient, iterative search, that typically shows greatly increased convergence speed compared to traditional approaches like design of experiments (DoE) or one‐factor‐at‐a‐time (OFAT), especially for complex systems [13, 14, 15, 16]. For readers seeking a broader introduction to BO in chemical reaction optimization, we refer to our recent review [17].

Bayesian optimization has already been demonstrated on both flow platforms [14, 16], where it is used mainly for continuous variables, and plate‐based HTE campaigns [13, 15] that mostly seek the optimal combination of discrete parameters. Recognizing the complementarity of these approaches—HTE for discrete variables, flow for continuous parameters—we envision a hybrid, two‐stage workflow where we leverage the strengths of both approaches. In the first stage, BO‐guided HTE is used to navigate a large search‐space encompassing a set of discrete parameters for a reaction of interest. In the second stage, the reaction is transferred to a flow reactor, where BO is used to optimize the continuous parameters around the best systems from the first stage.

This division in two stages presents several practical advantages. First, it leverages established HTE and flow technologies in a modular manner, thereby bypassing the need for complex flow systems capable of testing large numbers of categorical reagents, such as in robotic, multi‐step flow platforms [18]. Second, it avoids solubility and fouling issues that are detrimental to flow, but are easily handled by plate‐based HTE. Third, this approach is highly flexible and adaptable to different reactions, providing a template for efficient reaction optimization.

To validate this concept, a recently reported metallaphotoredox reaction for the formation of C(sp³)─C(sp²) bonds: the Amino Radical Transfer (ART) coupling was selected [19] (Scheme 1). These types of bond disconnections are particularly important in drug discovery, as they increase the fraction of sp³‐hybridized carbon in drug candidates on (hetero)aromatic scaffolds, thereby improving certain Absorption, Distribution, Metabolism, Excretion (ADME) properties [20]. The ART coupling demonstrates remarkable functional‐group tolerance, including unprotected alcohols, amines, and carboxylic acids, making it particularly relevant for organic synthesis and medicinal chemistry. It has also been recently applied to study the robustness of multiple parallel batch photoreactors and continuous flow setups, as well as its use in HTE optimization and subsequent Automated Parallel Synthesis [21]. Its oxygen and moisture insensitivity enhances operational simplicity and mitigates one source of potential outcome variability, critical to generate robust data.

SCHEME 1.

Amino Radical Transfer (ART) reaction.

Despite the advancements in ART chemistry, prior studies have predominantly focused on the use of aryl bromides as coupling partners, leaving other (pseudo)halide coupling partners underexplored. However, exploring these alternatives could prove advantageous, as synthetic campaigns often encounter scenarios where bromides are not readily available, are too labile, or remain unreactive. In this study, we aim to bridge this gap by carrying out our proposed two‐stage, BO‐driven approach for optimization across various (pseudo)halides to further generalize its application and potentially find divergent optimal reaction conditions. Additionally, we were interested to see whether this approach could result in the discovery of sustainable solvents and catalysts toward a potential implementation at scale.

In this study, we describe the two‐stage, BO‐guided optimization of this ART reaction for four aryl (pseudo)halides, combining HTE methodologies with our recently reported automated flow platform [22]. Specifically, we first optimized five discrete reaction parameters using HTE, after which the reaction was transferred to flow for further refinement of eight continuous variables. The campaign reveals distinct optima for all investigated aryl (pseudo)halides and, in all cases, uncovers solvent‐catalyst combinations with improved sustainability metrics relative to the literature‐reported benchmark. By demonstrating the efficacy of our hybrid approach on a challenging metallaphotoredox system, we seek to provide a generalizable blueprint for accelerated, sustainable process optimization across modern synthetic chemistry.

2. Results and Discussion

To evaluate the hybrid optimization workflow, the coupling between 4‐halo‐ or 4‐trifluoromethansulfonatebiphenyl 1 and benzylboronic acid pinacol ester (BPin) 2 were chosen as model substrates (Scheme 2). This pair allows all key species (starting material, product) and expected impurities (i.e., dehalogenated arene and hydrolyzed BPin) to be monitored via UPLC‐PDA through the presence of an aromatic chromophore (see Section S1.4). Although Suzuki‐Miyaura cross‐couplings with aryl triflates and non‐activated aryl chlorides are known, an ART‐like reaction with these has, to the best of the author's knowledge, not been reported yet.

SCHEME 2.

ART model reaction. Optimization parameters for the two stages are indicated. The four halides do not represent a variable, but rather four parallel campaigns.

Optimizing a metallaphotoredox reaction is expected to be highly complex, as both the photochemical and nickel cycles must be matched to each other, with lots of interplay between these cycles and the numerous parameters that govern them. For example, the choice of amine is expected to be heavily correlated to the photocatalyst, as the reported reaction mechanism (Scheme 3) requires a match between their redox potentials [19]. Likewise, oxidative addition of the chloride and triflate substrate to the active Ni complex is significantly slower than for the bromide and iodide, likely making that step rate‐limiting for these less‐reactive substrates [23]. Lower reactivity for chloride substrates for the ART reaction has been reported previously, although only nitrogen‐containing heteroaryl chlorides were considered [19]. Higher temperatures might mitigate this barrier and accelerate the nickel cycle [24] but may also decrease the stability of the generated alkyl radical [25], illustrating the risk of moving the two cycles out of sync. Additionally, Gisbertz et al. reported that for a metallaphotoredox C‐N coupling, matching the rate of photocatalyst‐mediated reductive elimination to that of the oxidative addition of the aryl halide to the active nickel complex is crucial to prevent nickel catalyst deactivation [26]. Taken together, these observations suggest that optimal reaction conditions might diverge across leaving groups, potentially enabling orthogonal functionalization in complex molecules bearing different (pseudo)halides.

SCHEME 3.

Mechanistic proposal for the ART reaction. Reprinted with permission from Speckmeier and Maier [19]. Copyright 2022 American Chemical Society.

2.1. Optimization of Discrete Reaction Variables with BO Guided HTE

We began by optimizing the five discrete reaction parameters using HTE combined with Bayesian optimization. Guided by previous reports on ART reactions expanded by our own experience in process development, a broad parameter space was defined, comprising twelve solvents, ten Ni ligands, five Ni precursors, ten photocatalysts (PC) and five amines, resulting in 30,000 possible combinations (Table 1). The parameter space was deliberately kept broad, including photocatalysts with redox potentials unlikely to be effective, to avoid biasing the optimization. In addition, this enabled the assessment of BO in a search space where not every combination is a hit. From a pragmatic point of view, this also allowed the use of standard ‘kits,’ rather than spending time curating tailored reagent sets.

TABLE 1.

Parameter space for the discrete variable HTE optimization campaign for the four aryl (pseudo)halides.

No.	Photocatalyst	Ered |*Ered | Eox | *Eox [V] a , b	No.	Nickel ligand
1	(Ir[dF(CF3)ppy]2(dtbpy))PF6	−2.01 | +0.89 | +1.69 | −1.21 [27]	1	4,4′‐Dimethoxy‐2,2′‐bipyridine
2	(Ir[dF(CF3)ppy]2(bpy))PF6	−1.37 | +1.32 | +1.69 | −1.00 [27]	2	1,10‐Phenanthroline
3	4CzIPN c	−1.24 | +1.43 | +1.49 | −1.18 [28]	3	Bathophenanthroline
4	4DPAIPN d	−1.52 | +1.10 | +1.34 | −1.28 [29]	4	4,4′‐Di‐tert‐butyl‐2,2′‐dipyridyl
5	3CzClIPN e	−1.16 | +1.56 | +1.79 | −0.93 [28]	5	4,4′‐Bis(trifluoromethyl)‐2,2′‐bipyridine
6	PheMesAcri.BF4	−0.50 | +2.17 | n.r. f | n.r. f [30]	6	N,N′‐Dimethylethylenediamine
7	PheMe2MesAcri.BF4	−0.58 | +2.09 | n.r. f | n.r. f [30]	7	2‐(2‐Pyridyl)benzoxazole
8	Ru(bpy)3(PF6)2	−1.33 | +0.77 | +1.29 | −0.81 [27]	8	2‐Methyl1,10‐phenanthroline
9	[Ir(dF(Me)ppy)2(4,4′‐dCF3bpy)]PF6	n.r. | n.r. | n.r. | n.r.	9	1,10‐Phenanthroline‐5,6‐dione
10	Ru(dtbbpy)3(PF6)2	−1.33 | +0.77 | +1.29 | −0.81 [31]	10	No ligand

No.	Solvent		No.	Amine	E1/2(D•+/D) [V] b
1	DMF		1	Pyrrolidine	+1.05 [19]
2	NBP		2	Piperidine	+1.07 [19]
3	Rhodiasolv Polarclean		3	Morpholine	+1.18 [19]
4	Acetonitrile		4	n‐butylamine	+1.35 [19]
5	Toluene		5	Cyclohexylamine	+1.39 g [32]
6	Propylene Carbonate		No.	Nickel precursor
7	Ethanol		1	Ni(PPh3)2Cl2
8	Teramyl alcohol		2	Ni(DME)Cl2
9	Anisole		3	Ni(TMHD)2
10	Ethyl acetate		4	Ni(MeCN)Cl2
11	Isopropyl alcohol		5	NiOTf2
12	2‐MeTHF

^a E^red = E_1/2(PC/PC^–), ^*E^red = E_1/2(PC*/PC^–), E^ox = E_1/2(PC^•+/PC), ^*E^ox = E_1/2(PC^•+/PC*).

^b all potentials measured in V (SCE) in MeCN unless indicated otherwise.

^c 4CzIPN = 1,2,3,5‐tetrakis(carbazol‐9‐yl)‐4,6‐dicyanobenzene.

^d 4DPAIPN = 1,3‐dicyano‐2,4,5,6‐tetrakis(diphenylamino)‐benzene.

^e 3CzClIPN = 2,4,6‐tri(9H‐carbazol‐9‐yl)‐5‐chloroisophtalonitrile.

^f Oxidation to the dication is typically irreversible and not tabulated for acridiniums.

^g Measured in DMF.

Because leaving‐group reactivity differs significantly, we ran four independent BO campaigns, one for each aryl (pseudo)halide, rather than treating the leaving group as an additional categorical variable. Otherwise, the algorithm might be biased toward the more reactive substrates, leaving less reactive ones underexplored. To maximize throughput, the four campaigns were run in parallel on a single 96‐well plate using an HTE workflow whose reproducibility was assessed (see Sections S2.1–S2.4). Each campaign was initialized with the same 24‐experiment Latin Hypercube Sampling (LHS) design, a space‐filling strategy that distributes the initial experiments broadly across the parameter space. Continuous parameters were fixed to literature‐reported defaults at this stage (2 eq. BPin, 1.5 eq. amine, 1% PC, 5% Ni, 0.2 M, 90 min) [19] so that we could leverage the core strength of the HTE platform: parallel evaluation of discrete reaction parameters. This is a deliberate trade‐off: although some interactions between discrete and continuous variables may remain unresolved in this first stage, it keeps the optimization experimentally tractable while still providing a chemically grounded basis for ranking discrete reaction systems.

In this first stage, yield was treated as the sole optimization objective. Multi‐objective optimization was not applied, because the goal was rapid identification of viable reaction systems in a very large, sparsely sampled search space. Introducing sustainability as a second formal objective would likely have reduced the efficiency of identifying productive systems by allocating experiments to conditions that may be preferable from a sustainability perspective, but are not chemically viable.

After the LHS initialization set, the BO algorithms proposed 20 new experiments per substrate; the remaining four wells duplicated the current best condition for each of the four aryl (pseudo)halides to track reproducibility and enable cross‐campaign information transfer. Immediately, the four campaigns started diverging, with different regions of the search space being explored. It was also clear, as expected, that the four substrates could be divided into two groups based on reactivity. The bromide and iodide showed quite high reactivity, while the chloride and triflate exhibited low yields. Progression plots of the campaigns are provided in Figure 1.

FIGURE 1.

Optimization progression of the HTE campaigns. Mean (top) and maximum (bottom) yields obtained in each iteration, showing overall improvement relative to the initialization set. Reaction: ART coupling between biphenyl‐X (X = Br, Cl, I, or OTf) and benzyl‐BPin.

After three BO‐guided iterations, we observed that successive suggestions were delivered only incremental gains; the model was repeatedly sampling narrow subsets of parameters, while still retaining an exploratory character expected for a search space of this size. Although true convergence cannot be claimed without an exhaustive search, the performance at this stage was sufficient to initiate the second stage. More importantly, exhaustive convergence was not the objective of this first stage, which was designed for rapid and data‐efficient screening to identify promising discrete reaction systems before subsequent refinement of the continuous parameters in flow. Before doing so, we added a fifth, manually curated plate as a final iteration aimed at sampling combinations with clear process advantages (greener solvents, organophotocatalysts over precious metal catalysis) that the BO algorithm had not yet explored. Guided by marginal yield profiles for each categorical variable (see Section S2.5), we enumerated untested combinations of the highest‐scoring individual parameter levels, prioritizing scalability and sustainability over further yield maximization.

This curated plate should be seen as a complementary addition to BO. The fact that these combinations might not have been suggested by BO can be explained by two factors. First, at this stage only a small fraction of the combinatorial search space had been sampled, and several of these pairings might have been proposed had the campaign continued. Second, sustainability considerations were not encoded as an explicit objective in the optimization campaign. Therefore, while this method does not capture potential interactions between parameters by relying only on marginal effects, it allows for steering the optimization pragmatically, and any conclusions drawn should be interpreted in this light. The curated plate is highly exploitative by design, which is reflected in Figure 1: the average yield across the plate is high compared to earlier runs for all four substrates, but the maximum yield does not improve for three of the four substrates.

In total, all four substrates were evaluated over five plates, representing 120 out of 30,000 possible experiments per substrate (<0.5% of the space). The resulting hit lists provided a sound starting point for the continuous‐parameter optimization in flow. This result clearly shows the potential benefit of algorithmic optimization, as navigating a broad search space quickly and efficiently is crucial from an industrial perspective.

To visualize how the algorithm navigated the 30,000‐point design space and to gain more insight into the reaction, we used a stacked bar chart (Figure 2). The plot shows how the distribution of discrete parameter choices evolves from the initial screening, where all levels are sampled equally, through subsequent iterations, highlighting which levels the BO algorithm explores and hinting at the favored parameter values. For example, the triflate campaign shows a strong preference for solvent 1 (DMF), while the chloride campaign shows a slight preference for solvent 2 (NBP) in later iterations. In contrast, for some other variables, the broader spread across iterations suggests that multiple levels remain competitive. This is illustrated by the Ni ligand choices, where 9 ligands are still sampled in the third iteration of the chloride campaign, compared to only 3 for the iodide campaign. The full experimental datasets for the four HTE campaigns, together with additional visualizations and their interpretation are provided in Sections S2.6 and S2.7.

FIGURE 2.

Evolution of discrete parameter choices over iterations during the HTE campaign. For each substrate (column), stacked bars show the frequency with which each level (colors) was sampled over successive iterations of the HTE campaign. This layout highlights how BO progressively focuses on, or sometimes re‐explores, specific solvent, photocatalyst, Ni precursor, Ni ligand, and amine settings, while abandoning others. Levels of each parameter refer to Table 1.

Reaction: ART coupling between biphenyl‐X (X = Br, Cl, I, or OTf) and benzyl‐BPin.

The top five conditions for each of the four substrates are shown in Table 2. Interestingly, for three out of the four campaigns, the final curated plate (fifth iteration), did not improve upon the BO conditions, suggesting the approach to generate this final iteration misses some nuance concerning parameter interactions that were picked up by the surrogate model. Several broad tendencies for these top‐performing conditions are apparent.

TABLE 2.

Top five best performing conditions from the HTE optimization campaign for each substrate. Systems selected for the next stage are highlighted in blue.

Solvent	Photocatalyst	Ni precursor	Ni ligand	Amine	Yield [%]
Br a
DMF	4CzIPN	Ni(PPh3)2Cl2	4,4′‐Dimethoxy‐2,2′‐bpy	Morpholine	99%
NBP	(Ir[dF(CF3)ppy]2(bpy))PF6	Ni(PPh3)2Cl2	4,4′‐Dimethoxy‐2,2′‐bpy	Morpholine	95%
NBP	3CzClIPn	Ni(PPh3)2Cl2	1,10‐Phenanthroline	Piperidine	91%
NBP	4CzIPN	Ni(PPh3)2Cl2	4,4′‐Dimethoxy‐2,2′‐bpy	Morpholine	78%
DMF	(Ir[dF(CF3)ppy]2(bpy))PF6	NiOTf2	1,10‐Phenanthroline	Morpholine	77%
Anisole	4CzIPN	NiOTf2	4,4′‐Dimethoxy‐2,2′‐bpy	Morpholine	71%[a]
Cl
NBP	4CzIPN	Ni(PPh3)2Cl2	4,4′‐Dimethoxy‐2,2′‐bpy	Morpholine	23%
NBP	4CzIPN	Ni(PPh3)2Cl2	4,4′‐Di‐tert‐butyl‐2,2′‐bpy	Morpholine	13%
DMF	(Ir[dF(CF3)ppy]2(bpy))PF6	Ni(MeCN)Cl2	1,10‐Phenanthroline	Cyclohexylamine	12%
DMF	(Ir[dF(CF3)ppy]2(dtbpy))PF6	Ni(PPh3)2Cl2	4,4′‐Dimethoxy‐2,2′‐bpy	Piperidine	11%
DMF	(Ir[dF(CF3)ppy]2(dtbpy))PF6	Ni(DME)Cl2	Bathophenanthroline	Cyclohexylamine	11%
I
2‐MeTHF	(Ir[dF(CF3)ppy]2(bpy))PF6	Ni(DME)Cl2	4,4′‐Di‐tert‐butyl‐2,2′‐bpy	Morpholine	>100% b
2‐MeTHF	(Ir[dF(CF3)ppy]2(bpy))PF6	NiOTf2	4,4′‐Di‐tert‐butyl‐2,2′‐bpy	Morpholine	>100% b
Propylene carbonate	[Ir(dF(Me)ppy)2(4,4′‐dCF3bpy)]PF6	NiOTf2	4,4′‐Di‐tert‐butyl‐2,2′‐bpy	Morpholine	88%
2‐MeTHF	[Ir(dF(Me)ppy)2(4,4′‐dCF3bpy)]PF6	NiOTf2	4,4′‐Di‐tert‐butyl‐2,2′‐bpy	Morpholine	82%
anisole	4CzIPN	NiOTf2	4,4′‐Di‐tert‐butyl‐2,2′‐bpy	Morpholine	79%
OTf
DMF	3CzClIPn	NiOTf2	1,10‐Phenanthroline	Piperidine	16%
DMF	3CzClIPn	NiOTf2	1,10‐Phenanthroline	Piperidine	16%
DMF	(Ir[dF(CF3)ppy]2(dtbpy))PF6	Ni(DME)Cl2	1,10‐Phenanthroline	Morpholine	16%
DMF	4CzIPN	Ni(DME)Cl2	1,10‐Phenanthroline	Pyrrolidine	14%
DMF	3CzClIPn	Ni(DME)Cl2	1,10‐Phenanthroline	Piperidine	13%

Reaction conditions: 0.2 M aryl halide, 2 eq. BPin, 1.5 eq. amine, 1 mol% photocatalyst, 5 mol% Ni complex, 40°C, 90 min, 365 nm LED, 300 µL scale.

^a For the bromide, six conditions are given, as the sixth system was taken to the next stage out of chemical interest.

^b Yield above 100% impossible; likely due to experimental error during weighing. Reaction: ART coupling between biphenyl‐X (X = Br, Cl, I or OTf) and benzyl‐BPin.

2.1.1. Solvent

For the bromide and chloride campaigns, DMF features prominently among the best conditions, along with NBP. This is perhaps not surprising, as N‐butyl pyrrolidone (NBP) is often considered as a greener alternative to DMF due to its similar physicochemical properties [33]. However, this does not seem to be the case for the triflate, as the 20 best conditions (see Table S9) all share DMF as a solvent. For the iodide, we were surprised to see 2‐MeTHF feature prominently across the best conditions, since the original ART paper suggested a strong electron‐donating solvent facilitated the radical transfer to the boronic ester [19]. Upon revisiting the data, we found that for the bromide an experiment with 2‐MeTHF as the solvent resulted in 48% yield, indicating that an ether‐type solvent is also able to work for that substrate as well. Finally, the presence of anisole for both the bromide (the sixth highest yield but included due to the use of anisole) and the iodide was an exciting prospect, as it ranks very favorably in industrial solvent selection guides [34, 35].

2.1.2. Photocatalyst

Across all four substrates, all of the three iridium catalysts and two of the three cyanoarenes (4CzIPN and 3CzClIPN) proved to perform well, with one of the latter coming out on top for all substrates except for the iodide.

2.1.3. Ni Precatalyst

Both Ni(PPh₃)₂Cl₂ and Ni(OTf)₂ dominate the best performing conditions across all four substrates, although all tested Ni precatalysts but one (Ni(TMHD)₂) feature at least once in the top five of any substrate.

2.1.4. Ligand

Interestingly, only four ligands are featured in Table 2: the two bipyridyl ligands and the two phenanthroline‐based ligands. 4‐4′‐Dimethoxy‐2‐2′‐bpy (MeO‐bpy) is common to the best performing conditions for both the bromide and chloride, while 4,4′‐Di‐tert‐butyl‐2,2′‐bpy (tBu‐bpy) is the only ligand in the top five conditions for the iodide. Notably, 1,10‐phenantroline dominates for the triflate, as not only the top five, but the top ten and 16 of the top 20 best‐performing conditions (see Table S9) share this ligand. Despite the algorithm having no prior information, this confirms recent studies investigating oxidative addition of aryl triflates to Ni complexes, where 1,10‐phenantroline was also identified as a high‐performing ligand [36, 37], underscoring the ability of data‐driven optimization to recover established reactivity patterns autonomously.

2.1.5. Amine

Morpholine is preferred for three substrates; for the triflate, piperidine (E_ox = 1.07 V [19]) is selected instead, possibly related to a higher quenching rate constant with the photocatalyst (3CzClIPN).

To select the systems to advance into the second stage, two pragmatic filters were applied:

1. Avoid DMF when performance allows it; instead, select a greener solvent.

2. Prioritize organophotocatalysts over Ir/Ru complexes.

Where those criteria conflicted, a judgment call was made in favor of scalability and cost rather than absolute performance. The following systems were taken forward, highlighted in Table 2.

Bromide For the bromide, two sets of conditions were retained: (i) NBP/4CzIPN/Ni(PPh₃)₂/MeO‐bpy/morpholine, and (ii) anisole/4CzIPN/Ni(OTf)₂/MeO‐bpy/morpholine. The first system was selected over the slightly better‐performing system with 3CzClIPN as a photocatalyst because of the significant cost difference. The second system, with anisole, was selected for the very attractive sustainability profile of the solvent.

Chloride The same NBP/4CzIPN/Ni(PPh₃)₂/MeO‐bpy/morpholine system as for the bromide led the table and was advanced directly.

Iodide A choice was required between anisole/4CzIPN/Ni(PPh₃)₂/tBu‐bpy/morpholine (organophotocatalyst, more sustainable solvent) and 2‐MeTHF/(Ir[dF(CF₃)ppy]₂(bpy))PF₆/Ni(DME)Cl₂/tBu‐bpy/morpholine (significantly higher yield). Both were retained for validation.

Triflate All top entries relied on DMF; the leading set (DMF/3CzClIPN/Ni(OTf)₂/1,10‐phenanthroline/piperidine) was selected.

2.2. Optimization of Continuous Variables on an Automated Flow Platform

2.2.1. Multi‐Objective Campaign for the Bromide Substrate

When targeting only yield, the BO algorithm has no incentive to use low catalyst loadings or limit the residence time to increase throughput. To prevent this, the bromide optimization campaign was treated as a two‐objective problem. Alongside yield, we introduced a desirability metric, that rewards low reagent cost and high productivity. A simplified definition is presented in Equation (1) (g∙day⁻¹∙€⁻¹; full equation in Section S4.3).

$$desirability=\frac{Productivity\left[g·da{y}^{-1}\right]}{Cost\left[€\right]}$$ (1)

A similar approach has recently been reported by Wagner et al., where the authors introduced a custom scoring function to penalize less sustainable conditions [38]. In contrast to that work, we carried out a true multi‐objective optimization, better visualizing the trade‐off present between these two objectives. To further prioritize yield, an option might be to introduce weights in the desirability objective function or even introduce a penalty for yield below a certain threshold.

The parameter space consisted of residence time [0.5–10 min], temperature [15–90°C], concentration [0.05–0.1 M], BPin equivalents [1–2], amine equivalents [1–2], Ni loading [1–10 mol%], photocatalyst loading [0.5–2 mol%] and light intensity [15%–100%]. The design space was deliberately chosen to be broad, to avoid biasing the optimization. At the same time, they were guided by pragmatic considerations: excessively high catalyst loadings, amine equivalents, or boronic ester equivalents were avoided, both to reflect realistic synthetic practice and to prevent unnecessary reagent consumption. Intermediate‐scale validation, details on the flow platform, and practical implementation details are provided in Sections S3 and S4.1–S4.5. During optimization, the campaigns were monitored, and the ranges could have been expanded if the algorithm converged toward their limits. The surrogate model was initialized using an LHS design of 15 points; thereafter, the algorithm proposed two experiments per iteration for ten iterations, to satisfy the total budget of 35 experiments. A benchmark condition was re‐run every third experiment to confirm stock solution stability and setup reproducibility (yield = 59.6 ± 2.1% over 11 experiments; see Section S4.6). A clear Pareto front emerged (Figure 3), representing the set of conditions for which improving one objective would require sacrificing the other, thus highlighting the trade‐off between them. The maximum yield obtained was 93%, and the maximum desirability was 6.04 g day⁻¹€⁻¹. The progress of the optimization and the optimal conditions for both objectives are shown in Figures 4, 5, and 6, illustrating steady improvement for both yield and desirability. The Pareto front also steadily expanded; of the last seven experiments, four are part of the Pareto front. The maximum yield in the initialization set was already high, so the absolute improvement was not very large, but the desirability tripled over the duration of the 25 algorithm‐suggested experiments. A more detailed analysis of trends based on parallel coordinate plots is provided in Section S4.7.

FIGURE 3.

Bromide Pareto front. Results of the optimization campaign for the ART reaction with the aryl bromide substrate. 15 initial LHS points were collected, after which the BO algorithm iteratively proposed 25 additional experiments. The figure illustrates the trade‐off between product yield and desirability: for the four points on the Pareto front, one objective cannot be improved without worsening the other. Reaction: ART coupling between biphenyl‐Br and benzyl‐BPin.

FIGURE 4.

Bromide optimization progress. Progression of the optimization campaign for the ART reaction with the aryl bromide. The blue line and points show the desirability for each experiment, while the green curve shows the maximum desirability up to that point, illustrating how the optimization campaign progressively identifies more favorable conditions. The plot also shows that 4 of the last 7 experiments lie on the Pareto front, highlighting continued improvement late in the campaign. The conditions resulting in maximal desirability are indicated on the right. Reaction: ART coupling between biphenyl‐Br and benzyl‐BPin.

FIGURE 5.

Yield optimization progress of the four aryl (pseudo)halides. Progression of the optimization campaigns of the ART reaction for all four substrates. The blue line and points show the obtained yield of each experiment, while the green curve shows the maximum obtained yield up to that point. The differing trajectories illustrate that both the extent and rate of improvement are strongly substrate‐dependent, consistent with the distinct optima ultimately identified for the four leaving groups.

Reaction: ART coupling between biphenyl‐X (X = Br, Cl, I, or OTf) and benzyl‐BPin.

FIGURE 6.

Optimal conditions for yield of the four aryl (pseudo)halides. Conditions maximizing yield for each of the four substrates, highlighting that the optimization converges to clearly distinct optima rather than to a broadly transferable recipe. Reaction: ART coupling between biphenyl‐X (X = Br, Cl, I, or OTf) and benzyl‐BPin.

2.2.2. Single‐Objective Campaigns for Chloride, Iodide, and Triflate

For the remaining three substrates, only yield was targeted, as this still remains the dominant performance metric. Experimental budget was set to 25 experiments per campaign. Although this budget is modest for an eight‐variable space, the objective here is to demonstrate the ability of the workflow to quickly generate results, which is crucial for triage of candidate routes, rather than exhaustive optimization. The same parameter space as for the bromide was used. Optimization progress and identified optima for all three substrates are presented in Figures 5 and 6, respectively.

Chloride: Because oxidative addition of aryl chlorides to the nickel complex is expected to be slower, as discussed earlier, the residence time boundaries were shifted to [10–45 min]. After an initialization set of 15 experiments, the algorithm shifted quickly toward higher temperatures to arrive at 35% yield over the next 10 iterations, an improvement over the results obtained using the literature‐reported continuous conditions in HTE (23%).

Iodide: Both the Ni complex and photocatalyst proved insoluble in 2‐MeTHF at the required concentrations for the stock solutions and were dissolved in DMF instead. Convergence was rapid: 97% yield was achieved after only 20 total experiments, and subsequent experiments offered no further improvement. To probe whether the presence of DMF in the reaction mixture had influenced this outcome, we prepared a new stock solution containing all reaction components in 2‐MeTHF and repeated the optimal conditions; the yield again reached 97%. This control suggests that DMF was introduced solely to address the solubility constraints of the stock solutions needed for the automatic flow platform, and did not materially affect the reaction performance.

Triflate: Solubility issues of the Ni(OTf)₂ complex in DMF necessitated a solvent screening to enable its use in a stock solution (see S4.5). Ultimately, 2‐methoxyethanol was identified as a solvent capable of dissolving the complex at sufficient concentrations. To make sure this solvent did not coordinate to the Ni complex and potentially alter its reactivity, we first performed a batch experiment where we added the Ni complex from a stock solution in 2‐methoxyethanol, affording a similar yield (16.8%) to the reaction in DMF (16.9%). We then repeated this experiment without ligand, using a stock solution of Ni(OTf)₂ in 2‐ methoxyethanol, and we obtained a very low yield (2% yield). Reassured, we proceeded with the optimization in flow. Even with extended residence time boundaries of [10–45 min], the triflate substrate proved even less reactive than the chloride, with the best identified conditions resulting in 28% yield. Additional control experiments probing the yield plateau are provided in Section S5. Nevertheless, this represents a valuable proof‐of‐concept for triflate coupling and can serve as a baseline for future exhaustive optimization.

The full flow campaigns and a more detailed analysis of these campaigns based on parallel coordinate plots are provided in Sections S4.7–S4.9.

2.3. Orthogonality of Identified Optima

The optimization campaigns yielded four markedly different “best” recipes, even for bromide and chloride, which share an identical catalyst–ligand–solvent system in the discrete space. To gauge how transferable these optima are, we ran a small cross‐screen in flow and summarized the results below.

Matrix 1 : changing only continuous variables (Table 3). For each substrate we kept its own discrete set (solvent, photocatalyst, Ni precursor, ligand, amine), but substituted the continuous optimum—residence time, temperature, concentrations, loadings—determined for each of the other substrates. Two observations emerge:

1. Minimal cross‐compatibility. Off‐diagonal yields fall sharply, for example, the bromide drops from 94% yield (self) to 64% under the optimal conditions for iodine, and the chloride drops from 35% to under 17% yield under any foreign setting.

2. Differences in reactivity range. Bromide tolerates chloride settings better (94%→82% yield) than chloride tolerates bromide settings (35%→17% yield).

TABLE 3.

Specificity of optima across substrates—only continuous variables. For each substrate, the optimal conditions of the continuous parameters of other substrates were applied while keeping the discrete parameters (the system) unchanged from the one that was identified for the substrate.

Only continuous parameters		Optimal conditions for
		Br	Cl	I	OTf
Performed reaction	Br	94%	82%	64%	20%	System: Br
	Cl	17%	35%	7%	4%	System: Cl
	I	53%	36% a	97%	2% a	System: I
	OTf	16%	16%	15%	28%	System: Otf

^a Reaction performed with a back pressure of 2 bar to prevent solvent boiling (2‐MeTHF).

Matrix 2: swapping the full optima (Table 4). Swapping both discrete and continuous parameters magnifies the divergence. Triflate is the most idiosyncratic, retaining only 3%–9% yield when forced into another substrate's recipe, underscoring its unique requirement for the specific catalytic system with phenanthroline ligand identified during the HTE optimization stage. Strikingly, forcing the bromide to the triflate's optimum all but inhibits the reaction, resulting in just 1% yield.

TABLE 4.

Specificity of optima across substrates—full optimum. For each substrate, the optimal conditions of both the discrete and the continuous parameters of other substrates were applied.

Full optimum		Optimal conditions for
		Br	Cl	I	OTf
Performed reaction	Br	94%	82%	37%	1%	System: Br
	Cl	17%	35%	3%	15%	System: Cl
	I	52%	57%	97%	46%	System: I
	OTf	9%	4%	3%	28%	System: Otf

Note: For the Br‐ and Cl‐substrates, the same optimal set of discrete variables was identified during the first optimization stage.

Taken together, the matrices show that substrate‐specific optimization is essential for this reaction: no condition is able to generalize across this small substrate scope. From a synthesis‐planning perspective, however, this orthogonality could be advantageous. As a preliminary proof‐of‐concept, methyl 3‐chloro‐5‐iodobenzoate 4 was first subjected to the optimal conditions identified for 4‐iodobiphenyl, affording methyl 3‐benzyl‐5‐chlorobenzoate 5 in 97% yield (Scheme 4). The crude reaction mixture was then subjected, without further optimization, to the optimal conditions identified for 4‐chlorobiphenyl, after addition of 1.12 and 1.21 equivalents of benzyl BPin and morpholine, respectively, resulting in 11% conversion to the desired doubly functionalized product 6. These preliminary results support the feasibility of exploiting substrate‐specific optima in a stepwise manner, although the second step was not further optimized and broader conclusions regarding generality remain outside the scope of the present work. Additional supporting data are provided in Section S6.

SCHEME 4.

Preliminary proof‐of‐concept for stepwise selective functionalization of a dihalogenated substrate guided by substrate‐specific optima. Application of the iodide‐optimal conditions to substrate 4 afforded 5 in 97% yield, and subsequent application of the chloride‐derived continuous conditions gave 11% conversion to product 6 without further optimization. Detailed reaction conditions and analytical data are provided in Section S6.

3. Conclusion

In this study, we have developed and demonstrated a two‐stage Bayesian optimization‐guided workflow, successfully applying it to the amino radical transfer (ART) coupling reaction between benzyl BPin and four distinct aryl (pseudo)halides. In the first stage, high‐throughput experimentation (HTE) efficiently screened a large space of 30,000 combinations, consisting of possible discrete parameter combinations—covering solvents, photocatalysts, nickel precursors, ligands, and amines. In the second stage, we transferred the optimal conditions to an automated flow platform and further optimized continuous variables to maximize yield and process viability.

By combining parallel exploration of discrete parameters in HTE with precise control of continuous parameters and process intensification in flow, we leveraged the complementary strengths of these two widely adopted technologies. For each substrate, 120 experiments in HTE over five iterations (<0.5% of the design space) and 25–35 experiments in flow resulted in tailored optimal values of thirteen parameters. This workflow extended the synthetic scope of the ART coupling to include aryl triflates, previously unexplored substrates, which represent an attractive addition to medicinal chemistry toolkits. Additionally, we enhanced sustainability and scalability by identifying viable replacements for the commonly used solvent DMF and precious metal‐based photocatalysts. Interestingly, the approach also proves able to identify literature‐reported reactivity trends without any prior knowledge.

The optimized conditions proved highly specific to each leaving group, demonstrating minimal transferability even between substrates sharing an identical catalytic system. This confirms that the leaving group exerts a dominant influence on reaction optimization, highlighting the value of substrate‐specific tuning for this reaction. At the same time, this orthogonality suggests interesting synthetic applications: sequential chemoselective functionalization of di‐hetero‐halogenated aryl scaffolds could potentially be achieved without resorting to protecting groups.

The approach highlights the importance of optimizing broadly in a high‐dimensional parameter space, which is made feasible here by utilizing standard, off‐the‐shelf laboratory equipment and capitalizing on the complementary strengths of HTE, automated flow technologies, and efficient optimization algorithms, such as Bayesian optimization. The complexity of the parameter space explored in this workflow, spanning multiple discrete and continuous variables, extends beyond the scope of previously reported optimization strategies, which typically either constrain continuous variables in HTE or limit discrete variable exploration in flow. The integration of Bayesian optimization enabled efficient sampling within a parameter space significantly broader than what traditional HTE methods could manage, as exhaustive full‐factorial screenings rapidly became impractical due to experimental and material constraints.

A key consideration is that decoupling discrete and continuous optimization stages may inadvertently leave some parameter combinations underexplored, potentially overlooking catalytic systems that are only active under conditions outside the initial HTE boundaries or not fully capturing interactions between discrete and continuous variables. This trade‐off was deliberately accepted, as a fully joint optimization over all discrete and continuous parameters would have led to an unacceptable increase in experimental complexity and burden. Nevertheless, this limitation could be further mitigated by iterative or integrated re‐optimization if needed for future reactions.

More broadly, this workflow is expected to be most useful for reaction optimization problems in which the search space is too large for exhaustive screening, while prior knowledge remains too limited to safely narrow that space in advance. Its value lies in allowing to retain a broad parameter space early in the optimization, before effort is concentrated on refinement of the most promising systems. In the present work, this staged strategy was implemented through HTE screening followed by flow optimization, but the broader concept is not inherently limited to that specific combination, or to laboratories with identical infrastructure. By contrast, for low‐dimensional problems or for reactions that are already well understood, the added value of introducing an algorithmic optimization layer may be less pronounced.

Ultimately, the reaction‐agnostic nature of this workflow positions it as a robust and practical tool for reaction development within our organization. Its demonstrated ability to rapidly deliver reliable conditions and provide a solid baseline for further optimization is expected to accelerate decision‐making and route assessment in future reaction development significantly.

Author Contributions

Stefan Desimpel: conceptualization, methodology, investigation, funding acquisition, writing – original draft, writing – review and editing, software, data curation, visualization. Jan Dijkmans: conceptualization, methodology, software, writing – review and editing, investigation, data curation. Florian Medina: methodology, writing – review and editing. Koen P. L. Kuijpers: methodology, writing – review and editing. Laurent Lefort: methodology, writing – review and editing. Santiago Cañellas: conceptualization, writing – review and editing. Kevin M. Van Geem: supervision, funding acquisition, writing – review and editing. Matthieu Dorbec: conceptualization, methodology, supervision, funding acquisition, project administration, writing – review and editing. Christian V. Stevens: supervision, funding acquisition, writing ‐ review and editing.

Conflicts of Interest

The authors declare no conflicts of interest

Supporting information

The Supporting Information contains detailed experimental procedures, BO implementation details, reproducibility analyses, additional campaign visualizations, and supporting analytical data. The authors have cited additional references within the Supporting Information [39, 40, 41, 42]. Supporting File 1: anie72680‐sup‐0001‐SuppMat.docx.

Data Availability Statement

The data that supports the findings of this study are available in the supplementary material of this article