by Nicholas A. Jose, Mikhail Kovalev, Eric Bradford, Artur M. Schweidtmann, Hua Chun Zeng and Alexei A. Lapkin
Abstract
Novel materials are the backbone of major technological advances. However, the development and wide-scale introduction of new materials, such as nanomaterials, is limited by three main factors—the expense of experiments, inefficiency of synthesis methods and complexity of scale-up. Reaching the kilogram scale is a hurdle that takes years of effort for many nanomaterials. We introduce an improved methodology for materials development, combining state-of-the-art techniques—multi-objective machine learning optimization, high yield microreactors and high throughput analysis. We demonstrate this approach through the optimization of ZnO nanoparticle synthesis, simultaneously targeting high yield and high antibacterial activity. In fewer than 100 experiments, we developed a 1 kg day−1 continuous synthesis for ZnO (with a space-time-yield of 62.4 kg day−1 m−3), having an antibacterial activity comparable to hydrothermally synthesized nano-ZnO and cetrimonium bromide. Following this, we provide insights into the mechanistic factors underlying the performance-yield tradeoffs of synthesis and highlight the need for benchmarking machine learning models with traditional chemical engineering methods. Methods for increasing model accuracy at steep pareto fronts, in this case at yields close to 1 kg per day, should also be improved. To project the next steps for process scale-up and the potential advantages of this methodology, we conduct a scalability analysis in comparison to conventional batch production methods, in which there is a significant reduction in degrees of freedom. The proposed method has the potential to significantly reduce experimental costs, increase process efficiency and enhance material performance, which culminate to form a new pathway for materials discovery.
Graphical Abstract
Introduction
Material innovation is a stepping-stone for technology development. Yet, development and commercialization of new materials is significantly limited by the expense, time and experience required. The typical time to bring a novel material to market is 10 to 20 years [1]. For nanomaterials, which are touted as next generation materials for many industries, developmental and production-related issues severely limit their commercial potential [2], [3], [4], [5], [6], [7]. Synthetic methods reported in literature are often too expensive or too hazardous to directly translate to the industrial scale. Key fundamental knowledge is also lacking. Recent studies have revealed complex relationships between material formation and mass transfer characteristics, such as hydrodynamics, which change significantly during scale-up [8]. Furthermore, commercialization requires the optimization of multiple competing criteria, such as cost and specific performance, which are often neglected in published research studies and patents. The target of creating an accelerated methodology for the development and mass-production of new materials has become especially urgent in times of increasing climate change, epidemics and economic instability. Several national efforts have already been initiated to tackle this challenge, including the Accelerated Materials Development for Manufacturing Programme (SG) [9] and Materials Genome Initiative (USA) [3]. In this work we present an accelerated methodology for materials development and scale-up, and demonstrate it through a scalable route to functional nano ZnO materials.
Materials development and scale-up requires an exhaustive amount of experimentation to understand the multivariable material-processing-property relationship. Scaling-up production from the laboratory (mg-g) to the pilot (kg-ton) and production scale (multi-ton) is often heuristic or empirical, amplifying the complexity and expense of development. Although mechanistic model-based scale-up is possible, accurate kinetic models of nanomaterial formation are both computationally expensive and difficult to derive. Several variations of larger equipment must be purchased, lab protocols must be re-evaluated, engineering parameters must be determined at each stage, and the labor required increases with each scale and experiment. While moving from each scale in this segregated, sequential fashion (i.e. the “stage-gate” approach) can lower risk, it involves large teams, which frequently lack proper information exchange [10].
Pilot-scale trials are the most critical step in scale-up, at which optimal engineering parameters for large-scale production are determined. Failures at the pilot scale are significantly more expensive than at the lab scale; work reverts to the laboratory and further investment in development is discouraged. Furthermore, the low availability of pilot production lines for nanomaterials, lack of industry technology readiness and poor knowledge of pilot processing amongst small-medium enterprises (SMEs) have recently been noted as barriers to the development of innovative material ecosystems [2].
Several tools have recently been developed to accelerate development. Coupling computational modelling with high-throughput experimentation can accelerate design and discovery [3], [11]. Machine-learning (ML) algorithms can increase the efficiency of data analysis and optimization [12], [13]. However, experimental applications of ML in materials optimization are typically focused on batch, mg-g scale synthesis. Conventional batch synthesis is not readily scalable because not all mass transfer parameters can be preserved when scaling to larger volumes [14].
Recently, the scalability of wet chemical synthesis has increased through the development of new processing techniques. Annular microreactors [15], spinning disk reactors [16], supercritical flow reactors [17] and helical flow reactors [18] can increase space-time-yield (STY – reaction yield per unit time per unit volume) by orders of magnitude while retaining control over nanoparticle size. Micromixers also provide precise control over mass transfer, which nanomaterials are sensitive to [19]. In contrast to conventional scaling of stirred tank reactors by increasing reactor volume (“scale-out”), micromixers are typically scaled by increasing the number of reactors in parallel (“number-up”) to conserve mass transfer characteristics [20]. In addition to the dimensionless parameters that have become the mainstay of scale-up methodology, hydrodynamic shear rate and residence time are also essential factors to consider in process intensification and scale-up of anisotropic nanoparticle production, for example, in the synthesis of layered double hydroxides [21], graphene [22] and titania nanotubes [23].
To approach the issues of scalability, efficiency and process complexity in nanomaterials development, a cross-disciplinary toolbox of acceleration techniques is needed. In this study we incorporated three tools: scalable processing technology, surrogate-based multi-objective optimization, and high-throughput testing. By doing this, we circumvented the classical stage-gate approach of product development, which is often upset by repeated failures and miscommunication between entities at different scales, and implement an “agile-inspired” development methodology, seen in Fig. 1.
To demonstrate the potential of the proposed approach in a case study, we developed a kg-per-day process for manufacture of highly active antimicrobial ZnO particles. ZnO possesses well-known antimicrobial properties, which stem from its surface activity, release of Zn2+ and catalyzed production of radical oxygen species [24], [25], which are correlated to its nanostructure [26], [27]. As a model system, ZnO possesses many of the challenges common to nanomaterial synthesis – morphological diversity, hard-to-scale published synthesis methods, and a complex performance-property relationship. Cost-effective antibacterial nanomaterials also have high social importance due to the rise of antibiotic resistance and high risk of surface-transmitted disease in public areas.
To synthesize ZnO in a scalable manner, we used annular microreactor synthesis (AMS), which was recently developed for the precise and high yield synthesis of two-dimensional materials, including layered double hydroxides [15] and metal-organic frameworks [28] with low clogging. Reagents and reactor conditions, including the shear rate and residence times, were varied to optimize antimicrobial efficiency and production efficiency. We employed the Thompson Sampling Efficient Multi-Objective algorithm (TSEMO); an approach for the simultaneous optimization of competing objectives with limited experimental evaluations [29], [30], [31]. Antimicrobial activity was assessed through the disk-diffusion test for inhibition of Escherichia coli growth, which allows a large number of samples to be tested in parallel. Mechanistic insights on ZnO synthesis with this approach were then drawn by characterizing a limited set of materials post-optimization. We then assessed this approach by comparing development time, safety, complexity and scalability to previously-reported continuous and batch processes.
Materials and methods
ZnO synthesis and yield
Synthesis of ZnO was conducted in an annular microreactor [15,28], which was assembled with three quartz capillary tubes from VitroCom (Tube 1 = 0.30 mm inner diameter × 0.4 mm outer diameter × 100 mm long, T2 = 0.50 mm inner diameter × 0.7 mm outer diameter × 100 mm long, and T3 = 1 mm inner diameter × 1.2 mm outer diameter × 100 mm long) in a “tube-in-tube” coaxial fashion. Stainless steel tee unions with 1/16” diameter tube compression fittings (Swagelok) with graphite ferrules (Restek) were used to connect the fittings stainless steel fittings and quartz tubes. A custom-built mount was used to precisely align the capillaries and fittings, which was evaluated visually using a portable microscope and magnifying glass. The length of the region in which reagents mix in the outermost tube was 50 mm.
A KDS Legato Dual Syringe Pump using disposable plastic 10 mL syringes was used to deliver liquid reagents to T2 and T3. The flow of filtered, compressed dried air through T1 was controlled using a Sierra SmartTrak C50L Mass Flow controller (20 L min− 1 max, 2% accuracy).
Reagent solutions of Zn-reagent (“A”) and alkaline reagent (“B”), prepared in the same solvent (either water or ethanol), were pumped simultaneously at equal flowrates into the outermost tubes (T2 and T3) of the annular microreactor while the compressed dried air flowed at high velocity through the innermost tube (T1). Solution A was pumped through T2 and solution B through T3. The resulting precipitates were centrifuged at 6000 rpm and rinsed three times in water or ethanol, ensuring that the final suspensions were of the same volume as their original reaction slurry. After rinsing, the solids content of the suspension and the corresponding dry-equivalent solid yield were determined gravimetrically by evaporating 1 mL of purified slurry in pre weighed glass vials at 110 ◦C. Three replicates were performed per experimental condition, using the average result for optimization. No unexpected or unusually high safety hazards were encountered.
Shear rates, pressure drops and velocities within the mixing region were calculated using the empirical model of Han et al. for wall stresses in gas-liquid annular flows for laminar and turbulent gas flows in tubes of 1 mm in diameter [32]. The averaged residence time (τR) was calculated using Eq. (1), where τR is the estimated average residence time (s), l is the length of the mixing region (m) and UL is the liquid film velocity (m s− 1).
The mean rate of energy dissipation per unit mass ε (m2s−3 or W⋅kg− 1) was calculated using Eq. (2), whereΔP is the change in pressure (Pa) and ρ is the liquid density (998 kg⋅m3).
The characteristic mixing time was then estimated from the relationship between the rate of energy dissipation and micromixing time for vortex engulfment [33], which is given by Eq. (3), where τE is the characteristic micromixing time (s) and υ is the kinematic viscosity (m2s− 1).
Disk-diffusion test for antimicrobial activity
In an adaptation of the Kirby-Bauer Disk Diffusion Test [34], Escherichia coli (ATCC 8739-BioRev) grown in Nutrient Broth (BioRev) at 37 ◦C was dispersed in 0.85% saline solution to an optical density of 0.1 at a wavelength of 600 nm. This dispersion was then spread on Mueller Hinton Agar (VWR) in petri dishes with sterile cotton swabs.
30 µL of 2.5 wt% ZnO suspensions were dropped onto disks of cellulose filter paper measuring 6 mm in diameter and dried. These disks were then placed face-down onto the inoculated plates, which were then incubated at 37 ◦C for 16–18 h. The diameter of the clear “inhibition” zone around each disk was measured. A ZnO control sample, which was known to reduce E.Coli colony forming units by > 99%, supplied by A*STAR SIMTech and synthesized according to reference [35], and a 2.5% solution of cetyltrimethylammonium bromide (CTAB – Merck), a known bactericide, were used as controls for each test. The average diameter of the control and CTAB were 9.9 ± 1.7 and 9 mm ± 0, respectively.
The antibacterial performance score, which represents the difference between the sample inhibition area and control inhibition areas, is given as S = DS − DC, where Ds is the sample inhibition zone diameter and DC is the inhibition zone diameter. For regions with no inhibition DS = 6mm, the diameter of the filter paper. Due to the variability of the method, three replicates were performed for each ZnO sample. Testing was done at a frequency of one batch (six samples) per day.
Experimental design and optimization
The experimental design methodology, shown schematically in Fig. 2, consists of the following steps:
1) Antibacterial performance S in the disk-diffusion agar method with E. Coli as test bacteria (in units of mm) and reactor time yield Y (in g of dry equivalent ZnO per minute) were selected as objectives.
2) 25 papers (references [25–27,35–57]) were surveyed for wet chemical precipitation methods that are compatible with annular microreactor synthesis to determine relevant synthesis variables. These synthesis variables were determined to be the zinc reagent anion (nitrate, sulfate, acetate or chloride), the alkaline reagent (NaOH or KOH), zinc and alkaline reagent concentrations and mixing intensity.
3) These variables were screened in a blocked factorial design to reduce the number of redundant variables and establish valid ranges on conditions for optimization, which amounted to 26 different synthesis conditions. From the results of these experiments, we saw that zinc reagent anions and alkaline reagent cations did not have significantly different effects on yield and performance. Zn(NO3)-6H2O (reagent A) and KOH (reagent B) were selected as reagents, and water was selected as the solvent due to its lower cost and less hazardous nature compared to most organic solvents. The four selected input variables and their ranges are summarized in Table 1.
4) Three iterations of the Thompson Sampling Efficient Multi Objective algorithm (TSEMO) were performed. An initial set of 20 experimental conditions was generated via Latin hypercube sampling (LHS) [58]. From this initial experimental dataset, TSEMO fits Gaussian process surrogate models (GPs) for each objective; from these surrogate models, the next set of experimental conditions that would best minimize model uncertainty and maximize the objectives (i.e. best approximate the Pareto front) is computed. After these conditions are experimentally tested, the optimization process is repeated until a specified maximum number of iterations has been reached. In this study, the covariances of the GP models were modelled by Mat´ern kernels of the 1/2 and 3/2 orders for yield and antibacterial score GPs respectively. For a more detailed description of TSEMO we recommend the reader to consult reference [29]. TSEMO code used for optimization was written in MATLAB and is available at [
https://github.com/Eric-Bradford/TS-EMO].
5) To further extend our approach, we have included another decision-making step—if the optimal conditions to reach the target objectives have not been determined, the process must revert back to step 2 and iterate. In this study, three iterations were used with 6 experimental conditions per iteration, were found to be sufficient. Hence, TSEMO chooses overall 18 experimental conditions to be carried-out.
6) To assess the Gaussian Process (GP) model quality, leave-one-out cross validation (LOO-CV) was performed, in which the model was trained on the experimental dataset 38 times, each time leaving one data point out for prediction [59]. To assess GP model predictions, we use the average absolute error (ε), which is defined in Eq. (4), where i is a sample point, ̂yi is the measured result at i, yGPi is the GP mean result and n is the number of samples. The errors of LOO-CV for yield and anti-bacterial score are referred to as εLOO-CV,Y and εLOO-CV,S.
7) Materials were synthesized at 6 chosen conditions with yield values of 0.6 g min− 1 and antibacterial scores ranging from − 0.6 to 3.8 mm to verify promising experimental conditions and to evaluate the model accuracy (i.e. “experimental evaluation”). Further, a limited set was further characterized with powder X-ray diffraction (XRD) and transmission electron microscopy (TEM). The errors of experimental evaluation for yield and antibacterial score are referred to as εexp,Y and εexp,S.
Materials and reagents
Reagent grade Zn(NO3)2-6H2O, Zn(SO4)-7H2O, Zn(Cl)2, Zn (CH3CO2), KOH (≥85%) and NaOH (≥98%) were obtained from SigmaAldrich. Deionized water (Millipore) and ethanol (96% – Singapore Chemical Reagent Co.) were used as solvents. E. Coli ATCC 8739, nutrient broth (HiMedia-MM244) and nutrient agar (HiMedia-MM012) were supplied by Bio-Rev. Whatman No. 5 filter paper (VWR), Petri dishes (90 × 14 mm), sterile swabs, culture tubes and sodium chloride (NORMAPUR analytical reagent) were supplied by VWR.
Powder X-ray diffraction
Suspensions were diluted in ethanol, drop-cast onto a non-reflective silicon wafer (100) and dried at 80 ◦C for 10 min. The powder x-ray diffraction pattern was collected with a Brucker D8 Advance Powder Diffractometer using Cu Kα radiation (λ = 1.5418 Å) at 40 kV from a 2θ of 3◦ to 70◦ with a step size of 0.02◦ and a scanning rate of 1.25◦ min− 1.
Transmission electron microscopy
Suspensions were diluted in ethanol, dropped on holey carbon 200 mesh copper TEM grids (InLab Supplies) and dried at ambient temperature. Images were taken with a JEOL 2100F FETEM at 200 kV.
Results and discussion
Optimization
From the 64 experiments performed (26 screening + 20 LHS + 18 TSEMO) an experimental Pareto front was resolved, ranging from anti-bacterial scores of − 1.7 to 5.2 mm and yields of 0.56 to 0.71 g min− 1 (shown in Fig. 3a). If we take an antimicrobial score > 0 mm as a lower bound specification and target maximum yield, we find that the highest performing experimental condition produces ZnO with a score of 2.17 mm and a yield of 0.70 g min− 1 (1.0 kg day− 1) in a single reactor.
Analyzing the set of conditions used (see Fig. 3b) we see that the initial LHS training set provides a sufficient spread of testing conditions. During subsequent TSEMO iterations, the experimental conditions narrow to the set of optimal conditions. Interestingly, CZn,A reaches a narrow region of optimal conditions after the first iteration, indicating that high concentrations can produce both high performance and high yield, which was not obvious from previous literature review. The results of each iteration are shown in Fig. 3a.
Fig. 3. TSEMO optimization, crossvalidation and experimental evaluation results. a) Antibacterial score and yields for each experimental iteration and the final corresponding Pareto fronts (modeled and experimental) b) Corresponding experimental conditions, where QL is the total liquid flowrate, QG is the gas flowrate, RKOH/Zn is the molar ratio of KOH to Zn, and Czn,A is the molar concentration of Zn(NO3) in reagent A. Data within the dashed lines are the results of LHC initialization and data within the solid lines are results of TSEMO optimization c) Modelled Pareto fronts across different TSEMO iterations and model targets for experimental evaluation. d) Model antibacterial scores and e) yields compared with measured yields and antibacterial scores at the same conditions, where the red dashed line is the ideal fit (100%) and error bars are model 95% confidence intervals. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
The modelled Pareto front (i.e. the Pareto front of the GP model shown in Fig. 3a) lies along the experimental Pareto front (i.e. the Pareto front of the experimental measurements). The modelled Pareto changes in shape as more data is added (seen in Fig. 3c) showing an increase in accuracy with each iteration. The surprising steepness of the Pareto front and the narrow window of optimal processing conditions illustrate the sensitivity of this tradeoff to processing conditions and highlight the importance of finely controlled process parameters in the synthesis of nanomaterials.
The results of model cross-validation and experimental evaluation are shown in Fig. 3d and e, where the GP model predictions are compared with the respective measurements. The greater 95% confidence intervals of modelled antibacterial scores reflect the larger variance in experimentation. εexp,S and εexp,Y were 2.3 mm and 0.08 g min− 1 respectively, while εLOO-CV,S and εLOO-CV,Y were 1.5 mm and 0.04 g min− 1. 89% of the cross-validation results lied within the 95% confidence interval of model predictions (seen in the error-bars of Fig. 3d and e), indicating the accuracy of the model. Within the experimental evaluation, 4/6 of the yields and 5/6 of the antibacterial scores lied within the 95% confidence interval of model predictions.
Model variance is strongly influenced by the precision of experimental measurements. Antimicrobial tests had an average standard deviation of 1.03 mm (9.8% of the measurement range). This is close to εLOO-CV,S and is likely due to variation in biological samples, filter papers and dosing of ZnO. Yield results had standard deviations of 0.04 g min− 1 (5.8% of the measurement range), and is the same as εLOO-CV,Y, and may be a result of uneven sampling and loss of sample during purification.
Cross-validation errors were likely lower than the errors of experimental evaluations due to the increased sample size. Discrepancies between the TSEMO model predictions and experimental results may be a result of several factors. Model deviation likely arises from experimental noise as well as the DOE selected. TSEMO selects experimental points with dual objectives increasing model accuracy (“exploration”) and optimizing outputs (“exploitation”) – which involves some sacrifice of global model accuracy. To increase the accuracy of the GP model, more data would be needed.
Furthermore, the larger deviations in predictions of antimicrobial activity may be attributed to the larger amounts of biological variation in samples, which in turn increase the error of the model. There may also be other variables that affect synthesis that are not accounted for in the model, for example, variations in the starting materials used across different batches from a single supplier and the microbiologist performing each test. Particle characteristics were also not considered as model parameters within the study. The robustness of model predictions should be honed in the future by conducting larger numbers of experiments and including more variables.
In an extension to this study, pairing model predictions with output targets can guide further development and scale-up trials. For example, if we target an antibacterial score of ≥ 0 mm, the modelled Pareto front can be used to predict promising process conditions with 95% certainty. Processing tolerances could also be incorporated for sensitivity analysis. For example, although yields of up to 0.7 g min− 1 can be achieved, the range of conditions that can achieve this may be very narrow, and a yield of 0.6 g min− 1 may be a safer experimental target (see Fig. 3c). The model yield can then be used to estimate the number of reactors needed for scale-up.
In further studies it is necessary to benchmark the GP model obtained from TSEMO to those using traditional chemical engineering methods, from simpler methods like empirical numerical models to the more complex, deterministic models that couple computational fluid dynamics, molecular dynamics and population balance models for crystal growth.
It is important to note the limitations of the specific methods used in our case study. Wet chemical synthesis and microreactors are not universally suitable for every new material. Process selection should initially be guided by practical knowledge; however, experimentalists still benefit from using efficient synthesis methods with well-defined engineering parameters early in the development. The design of experiment and/or statistical model used should also be tailored to the problem at hand. TSEMO is appropriate when multiple competing objectives exist, the variables used are continuous and experiments are expensive to evaluate. For problems in which objectives are non-competing, variables are discrete or large datasets are readily obtained, the experimental problem may be significantly different and the present methodology can be extended through the selection of another DOE approach [60].
Conclusions
In summary, the pairing of annular microreactor synthesis, the multiobjective optimization algorithm TSEMO and highthroughput testing for the development of antibacterial ZnO has yielded three significant results. An optimized process for 1 kg per day production of a material with activity comparable to a commercially available antimicrobial and conventionally synthesized nano-ZnO was developed in less than 100 experiments. A brief analysis of the materials synthesized in these trials suggested
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
This project was funded by the National Research Foundation (NRF), Prime Minister’s Office, Singapore, under its Campus for Research Excellence and Technological Enterprise (CREATE) program as a part of the Cambridge Centre for Advanced Research and Education in Singapore Ltd (CARES) and under the SMART Innovation Centre Grant (ING-000630).
References
There are more references available in the full text version of this article.
This article is cited by
Science Direct