Solvent screening method

Abstract

There is provided a method for monitoring of one or more process parameters of a chemical process to convert one or more input substances to one or more output substances at different solvent ratios, comprising (a) providing a plurality of input streams, each comprising a solvent or solvent mixture and one or more input substances for the chemical process, wherein at least two of the input streams comprise a distinct solvent or solvent mixture; (b) delivering one or more of the input streams to a flow reactor; (c) varying the ratio of solvents within the flow reactor by changing relative flow rates of the input streams to the flow reactor; and (d) monitoring one or more process parameters at different solvent ratios. There is provided a system for carrying out the method. There is also provided a method of training a machine learning model to predict one or more process parameters of the chemical process, and a method of designing and performing the chemical process.

Description

[0001] SOLVENT SCREENING METHOD

[0002] Field of the invention

[0003] The invention relates to a method for monitoring one or more process parameters of one or more chemical processes at different solvent ratios, and a system for carrying out the method. The invention also relates to a method of training a machine learning model to predict one or more process parameters of the chemical process, and a method of designing and performing the chemical process.

[0004] Background to the invention

[0005] Organic chemical reactions, and other chemical processes, require a set of conditions to run successfully which often are selected based on chemical intuition and precedent. Recently, modelling and more data driven approaches have been implemented in chemical industry. However, the rate of empirical data collection is severely limiting, exacerbated by a reliance on old technologies and a culture of status quo within large companies.

[0006] Solvent choice is a factor that has been particularly challenging to make progress on as it is hard to computationally model or collect diverse data on solvent effects on processes and reactions by traditional methods. There is a widespread desire to move towards benign, cheap, sustainable solvents. Yet, many reactions are still performed in hazardous environmentally detrimental solvents which are being regulated against or suboptimal solvent systems. Finding replacements for these solvents using solvent mixtures and collecting accurate scalable data on different solvation systems using already commercially available benign solvents is of pressing need to develop new chemical processes economically and efficiently.

[0007] Summary of the invention

[0008] In a first aspect there is provided a method for monitoring one or more process parameters of a chemical process to convert one or more input substances to one or more output substances at different solvent ratios, comprising: a) providing a plurality of input streams, each comprising a solvent or solvent mixture and one or more input substances for the chemical process, wherein at least two of the input streams comprise a distinct solvent or solvent mixture; b) delivering one or more of the input streams to a flow reactor; c) varying the ratio of solvents within the flow reactor by changing relative flow rates of the input streams to the flow reactor; and d) monitoring one or more process parameters at different solvent ratios. In a second aspect there is provided a method of training a machine learning model to predict one or more process parameters of a chemical process to convert one or more input substances to one or more output substances at different solvent ratios of at least two solvents or for a solvent, the method comprising: monitoring the one or more process parameters at different solvent ratios of one or more pairs of solvents using the method of any preceding claim; recording training data pairs of respective solvent ratio data points and associated process parameter data points, the solvent ratio data points comprising a value of solvent ratio and the process parameter data points comprising respective one or more values of the one or more process parameters; and training a machine learning model using the training data pairs to output a predicted process parameter data point in response to an input of a solvent ratio data point.

[0009] In a third aspect there is provided a method of designing a chemical process to convert one or more input substances to one or more output substances, the method comprising using a machine learning model trained according to the method according to the second aspect to predict respective one or process parameters for a plurality of ratios of solvents of a pair of solvents, or for a solvent; selecting a solvent ratio based on the predicted respective one or more process parameters; selecting the solvents of the pair of solvents at the selected solvent ratio, or a solvent, for the process.

[0010] In a fourth aspect there is provided a system for monitoring one or more process parameters of a chemical process to convert one or more input substances to one or more output substances at different solvent ratios, comprising:

[0011] (i) a plurality of variable flow pumps, each in fluid communication with an input stream comprising a solvent or solvent mixture and the one or more input substances for the chemical process, wherein at least two of the input streams comprise a distinct solvent or solvent mixture;

[0012] (ii) a flow reactor in fluid communication with each of the variable flow pumps; and

[0013] (iii) an analytical unit in fluid communication with the flow reactor, configured to monitor one or more process parameters.

[0014] Brief Description of the Drawings Below is a concise description of the accompanying figures, each of which illustrates various aspects of the invention disclosed herein. These figures are referenced throughout the specification to enhance understanding and clarity regarding aspects of the invention, but should not be considered limiting upon the scope of invention.

[0015] Figure 1 shows a schematic representation of the fluidic path and key components of an example flow system for carrying out the methods described herein.

[0016] Figure 2 shows an example chromatogram of a method using HPLC for monitoring a unimolecular reaction.

[0017] Figure 3 shows an example HPLC calibration curve collected via a flow rate ramp between a solution containing starting materials including 1d (Z = H) and a solution containing ethyl benzene as an internal standard.

[0018] Figure 4 shows a visualisation of the HPLC sample analysis for the peak areas of 1, 2, and 3 of example 1 plotted as yield vs. solvent mixture composition. The data series represented are for a bidirectional ramp in solvent composition to confirm that the data at each solvent composition is independent of the ramp direction (hysteresis).

[0019] Figure 5: An example chromatogram of the HPLC method for monitoring six reactions in a “one-pot” method. The lower diagram shows a diagrammatic representation of the HPLC method.

[0020] Figure 6 shows a visualisation of the HPLC sample analysis for the peak areas of 1a-f, 2a-f , and 3 of example 2, showing yield of each reaction vs. % water in ethanol.

[0021] Figure 7 shows an example chromatogram of the HPLC method for example 3. The lower two diagrams shows a diagrammatic representation of the HPLC pump method and the dilution pump method.

[0022] Figure 8 shows a visualisation of the HPLC sample analysis for the peak areas of 3, 6 and 7 of example 3, showing yield of the competing reactions vs. % water in ethanol.

[0023] Figure 9 shows a flow chart diagram illustrating a method of training a machine learning model to predict one or more process parameters of a chemical process. Figure 10 shows a flow chart diagram illustrating another method of training a machine learning model to predict one or more process parameters of a chemical process.

[0024] Figure 11 shows a flow chart diagram illustrating another method of training a machine learning model to predict one or more process parameters of a chemical process.

[0025] Figure 12 shows a flow chart diagram illustrating a method of designing a chemical process to convert one or more input substances to one or more output substances.

[0026] Figure 13 shows predicted values plotted against true values for a number of solvent ramp experiments from Example 4. Each fold was trained on all ramps except one, which is shown in the title and the plots show the machine learning model’s predictions on the unseen ramp. Shown is the uncertainty estimates for each predictions via ± 2 times the standard deviations, creating 95% credible intervals. The true values generally fall within the confidence intervals.

[0027] Figure 14 shows a machine learning model trained on all ramps except for water — methanol and HFIP — I PA, from Example 4. Shown is the model predictions on the unseen ramps. The uncertainty estimates can be used to choose which ramp to gather next.

[0028] Figure 15 shows the Mean Squared Error (MSE) of a machine learning model, against number of ramps trained on, from Example 4. On the right hand side the same results are shown but zoomed in. It is observed that by using an Ml-based selection criteria better modelling can be obtained.

[0029] Detailed Description

[0030] The method herein provides a way of monitoring one or more process parameters of one or more chemical processes to convert one or more input substances to one or more output substances at different solvent ratios.

[0031] In particular, the method herein provides, for the first time, a way of collecting large amounts of data in a continuous manner and in bulk. The output is a series of data points providing richer information on the performance of the process in different solvent systems. This data can then be used to quickly and efficiently select the most appropriate solvent system for the chemical process(es) and to inform further process decisions and technoeconomic analysis for solvent system choice. The data collected according to the methods herein is more accurate and reliable than data collected in batch because it prevents solvent evaporation at elevated temperatures by allowing safe performance of chemical processes under pressure and without headspace. There is also more homogenous heating across the solutions due to the large surface area to volume ratio. The method is also safer than carrying out the same steps in batch as it allows for a smaller reactor volume and requires less exposure to, and less use of, potentially hazardous chemical reagents and solvents.

[0032] Compared to the flow methods in the prior art for collecting this data, which are performed under steady state conditions, the methods herein advantageously provide a large increase in data collection efficiency and density.

[0033] T raditionally, a solution in the given solvent mixture would need to be primed in the flow system and then left under the set flow rate and temperature for roughly three reactor volumes (and residence times) to equilibrate. Therefore, if the residence time of the system is approximately 20 minutes, it would take ~60 minutes to collect one data point, not including priming time and solution make up time. In the methods described herein, it would be possible to collect 30 independent solvent mixture data points in 180 minutes, equivalent to one solvent mixture point every 6 minutes: this is a ten fold time and material saving (as the flow rate of reaction solution is constant over this time).

[0034] The method

[0035] The method allows for the monitoring of process parameters of a chemical process to convert one or more input substances to one or more output substances at different solvent ratios.

[0036] The method comprises a first step of (a) providing a plurality of input streams, each comprising a solvent or solvent mixture and one or more input substances for the chemical process, wherein at least two of the input streams comprise a distinct solvent or solvent mixture. The input streams may suitably be each contained within separate receptacles (e.g. Duran bottles fitted with inlet tubing, and optional in-line filters), each receptacle being connected to a means of conveying the stream.

[0037] By way of illustration, if two binary solvent mixtures of methanol and ethyl acetate are to be used in the method, there could be two input streams with the same input substances, but with a first input steam with 100% methanol as a solvent, and a second input stream with 100% ethyl acetate as solvent. The method then involves a second step of (b) delivering one or more of the input streams to a flow reactor. This can be carried out by using any suitable means for conveying the input streams to the flow reactor, such as by using a number of variable flow pumps (e.g. a Gilson 305 or 307 HPLC pump), each in fluid communication with each of the receptacles containing the input streams, and the flow reactor. The flow reactor can be any suitable reactor for the selected chemical process.

[0038] The method then comprises a step of (c) varying the ratio of solvents within the flow reactor by changing relative flow rates of the input streams to the flow reactor. If using variable flow pumps, this could be achieved by using a control system programmed to vary the flow rate from the input streams to the flow reactor.

[0039] For example, the method might suitably involve ramping the relative flow rates of the input streams to the reactor linearly so that the solvent passing through the flow reactor gradually switches from one solvent to the other over time. In the above illustrative example the solvent would linearly ramp from 100% methanol to 100% ethyl acetate, with all mixtures in between being observed during.

[0040] Varying the ratio of solvents may comprise ramping of the relative flow rates. It will be appreciated that ‘ramping’ refers to a gradual or progressive change in flow rate. A ramp may occur with a change in flow rate in a single direction or multiple directions, i.e. increasing the flow rate of an input stream, followed by decreasing the flow rate of that same input stream.

[0041] Preferably, varying the ratio of solvents comprises a continuous linear or non-linear ramping of the relative flow rates. Continuously should be understood to mean that the flow rate of the one or more input streams, at some point in the method, is changing continuously. A linear ramp means whenever a ramp is performed (j.e. when the system is not at a fixed solvent ratio) the ramp may follow a form according to the equation y = mx +c. For example, linear ramps in solvent ratio may go from 0 to 50% and then hold at 50% before going linearly back down to 0%. In contrast, a non-linear ramp does not follow the function y = mx +c, and may be a more complex function dictating the ratio of solvent flow rates, such as in the form of an exponential or quadratic function.

[0042] The constant total flow rate to the flow reactor may be changed or may be maintained at a constant rate throughout the method. However, preferably, in step (c) the ratio of solvents is varied whilst maintaining a constant total flow rate to the flow reactor. This may be achieved, for example, by using a control system that adjusts the flow rates of the input streams to the flow reactor in a way such that flow is maintained to the flow reactor at a constant rate.

[0043] It is possible for the solvent ratio to be varied at the same time as changes in residence time or changes temperature. However, it is preferred that the residence time and temperature of the flow reactor are kept constant throughout the chemical process or reaction.

[0044] The method then involves (d) monitoring one or more process parameters at different solvent ratios. The method of monitoring chosen will depend on the chemical process(s) that is(are) selected in the method, but may include the use of analytical tools such as on-line HPLC that measures the level of the input and / or output substances, which in turn can be used to determine other performance parameters related to the chemical process(es).

[0045] The output of the above method is a series of data points for the one or more process parameters of the chemical process across a spectrum of the ratios of solvent or solvent mixtures. The method allows, for the first time, for this data to be collected in a continuous manner and in bulk.

[0046] There are a number of advantages associated with this method. The output provides richer information on the performance of the process in a variety of different solvent systems. A particular advantage is that this data can then be used to quickly and efficiently select the most appropriate solvent system for the chemical process and to inform further process decisions and technoeconomic analysis for solvent system choice.

[0047] Each data point in these series would previously have required an individual experiment to be performed either by batch or flow methods. Inherently the data collected according to the method herein is more accurate and reliable than the same data collected in batch. It is more accurate because it prevents solvent evaporation at elevated temperatures by allowing performance of chemical processes under pressure and without headspace. There is also more homogenous heating across the solutions due to the large surface area to volume ratio. The method is also safer than carrying out the same steps in batch as it allows for a smaller reactor volume and requires less exposure to, and less use of, potentially hazardous chemical reagents and solvents. Chemical processes

[0048] In the method the chemical process (or chemical processes if multiple) will be selected by the user according to their needs. It should be recognised that the method described herein is not limited to any particular type of chemical processes.

[0049] It will be appreciated that the method herein may be for concurrently monitoring one or more process parameters of multiple chemical processes concurrently. Any discussion of features relating to a method of monitoring a chemical process herein is self-evidently applicable to embodiments wherein multiple chemical processes monitored, e.g. in a “one-pot” method.

[0050] The chemical process(es) may be selected from a chemical reaction, an extraction, a crystallisation, and a quenching step.

[0051] Chemical reactions

[0052] Preferably, the chemical process is a chemical reaction to produce a product from one or more starting materials, wherein the one or more input substances are one or more starting materials and the one or more output substances are the one or more products. If the chemical process is a reaction then the method may comprise a step of determining reaction performance from the monitored one or more process parameters at different solvent ratios.

[0053] The chemical reaction(s) may be a unimolecular, bimolecular or multicomponent chemical reaction, or any combination thereof. It will be appreciated that the method could be applied to reactions with more than two starting materials or more than two products.

[0054] Non-limiting examples of unimolecular reactions include isomerization, deprotonation of acids, decarboxylation, first-order elimination (E1), thermolysis, and intramolecular cyclization. Nonlimiting examples of bimolecular reactions include SN2 reactions, E2 elimination, hydrolysis of esters, amide bond synthesis reactions, Grignard reaction, nucleophilic aryl substitution, Diels- Alder reaction, Michael addition, and Wittig reaction. Non-limiting examples of multicomponent reactions include the llgi reaction, Passerini reaction, Biginelli reaction, Mannich reaction, Kabachnik-Fields reaction, and Strecker synthesis.

[0055] The chemical process may be a chemical reaction selected from a homogeneous organic chemical reaction, a heterogenous catalysis reaction, a homogeneous catalysis reaction, a biphasic liquid-gas reaction, a nano particle synthesis reaction, a crystallisation reaction, a biological reaction, a fermentation reaction, a polymerisation reaction. If the method involves multiple chemical processes, then the chemical processes may be any combination of these reactions.

[0056] The chemical process(es) may be a heterogeneous catalysis reaction (solid-liquid), including but not limited to: hydrogenation of alkenes, Friedel-Crafts alkylation, selective oxidation of alcohols, and photocatalytic degradation of pollutants. The flow reactor may be fitted with a solid catalyst, and the one or more input substances may be liquid starting materials that react with the catalyst to produce one or more output substances that are the product(s).

[0057] The chemical process(es) may be a homogeneous catalysis reaction, including but not limited to: hydroformylation, Heck reaction, Sonogashira coupling, Suzuki coupling, asymmetric hydrogenation, and olefin metathesis. The one or more input substances may be liquid starting materials and liquid catalyst that react in the flow reactor to produce one or more output substances that are the product(s).

[0058] The chemical process(es) may be a biphasic liquid-gas reaction, including but not limited to hydroformylation of olefins, aerobic oxidation of alcohols, and carbonylation of methanol to acetic acid, or any combinations thereof.

[0059] It is particularly preferred that the chemical process(es) is(are) a homogeneous organic chemical reaction between multiple reactants with or without a catalyst.

[0060] The chemical process(es) may not necessarily involve any molecular transformation. For instance, the chemical process(es) may be a process of preparing a formulation (e.g. mixing components and then measuring parameters to indicate formulation stability, turbidity, coagulation, etc.), or a process for cleaning a reactor (e.g. by passing cleaning chemicals through a flow reactor and measuring the removal of impurities). The chemical process(es) may be regeneration of a catalyst (e.g. of a fresh or spent catalyst before use in a chemical reaction), or a process for extraction of chemical materials from biomass or recycled plastic waste. It will be appreciated that the same analytical units and equipment set up as described herein for monitoring a chemical reaction may be used for such chemical processes.

[0061] Input substances and output substances

[0062] It will be appreciated that, depending on the chemical process(es) chosen, the input substance(s) and output substance(s) will be selected accordingly by one skilled in the art. Any number of input and output substances may be involved depending on the chemical process chosen. Similarly, the invention is not limited to two input streams and it will be appreciated that many input streams may be used with the method, wherein each has a distinct solvent or solvent mixture. This is advantageous as it allows for a large number of solvent systems to be screened in the method at once, allowing more efficient collection of data.

[0063] The plurality of input streams may preferably be three or more input streams, wherein three of the input streams comprise a distinct solvent or solvent mixture. The plurality of input streams may be six or more input streams, wherein six or more of the input streams comprise a distinct solvent or solvent mixture.

[0064] It will be appreciated that it is possible to vary other reaction conditions at the same time as varying the solvent ratio. For example, if a compound has a low solubility in a certain solvent it may be preferable to also have the overall reaction concentration to be varying at the same time as the solvent composition. In such instances the plurality of input streams may comprise the input substance(s) in different concentrations.

[0065] However, it is preferred that the plurality of input steams each comprise the input substance(s) at the same concentration (e.g. the same reactant stoichiometry) in the solvent or solvent mixtures.

[0066] Flow reactor

[0067] The term “flow reactor” is used herein to refer to the region wherein the chemical process takes place. It will be appreciated that for some chemical processes, the flow reactor may simply be represented by tubing (e.g. passing through an oven to induce the chemical process) before it reaches an analytical tool (e.g. HPLC), rather than a separate reactor component.

[0068] Any suitable flow reactor known in the art may be used so long as it is suitable for the chemical process selected for monitoring.

[0069] The flow reactor is suitably selected to possess means to induce the chemical process, or suitably connected to a means to induce the chemical process, for example, through heating, cooling, microwave-activation, pH-change, or photo-activation, and the like. However, it will be appreciated that, for some chemical processes, induction may occur by the mixing of the input streams themselves. If the chemical process is a biphasic liquid-gas process, then preferably the flow reactor is a gas-liquid contactor, preferably a tube-in-tube reactor, bubble type reactor, or segmented flow system. The reactor may also be a Continuous Stirred-Tank Reactor (CSTR).

[0070] The reactor is preferably a tubular reactor or Continuous Stirred-Tank Reactor (CSTR) for homogeneous reactions initiated by heating, cooling, microwave-activated, photo-activated, or simply initiated through mixing (or a combination of the same).

[0071] For a heterogeneous process or reaction (involving liquid-solid, or liquid-solid-gas), the reactor is preferably a packed bed reactor or Continuous Stirred-Tank Reactor (CSTR).

[0072] Most preferably, the flow reactor is a thermally heated tubular reactor, e.g. with fluid dynamics within the assumptions of an ideal plug flow reactor.

[0073] Solvents

[0074] The method of the invention may be carried out using any suitable solvent or solvent mixture for the chemical process(es) chosen.

[0075] The solvent or solvent mixture may be a liquid or a gas, but preferably is a liquid at room temperature and pressure.

[0076] The solvent or solvent mixtures may be selected depending on the chemical process(es) to be monitored. Organic solvents are preferred, and it is particularly preferable these are liquid at room temperature. Suitable organic solvents include polar, apolar, protic, aprotic, halogenated, aqueous and aromatic solvents.

[0077] The solvent or solvent mixture may comprise a solvent selected from deep eutectic solvents and other inorganic derived solvents. It will be appreciated that specialised pumps or extruders may be needed to implement the method using these types of solvents, and such equipment and its implementation would be apparent to one skilled in the art.

[0078] The solvent or solvent mixture may comprise one or more supercritical fluids, including those which are gaseous at room temperature and atmospheric pressures. Preferably, the supercritical fluid is supercritical CO2. To implement a method as described herein whereby the solvent is supercritical CO2, then a setup similar to that used in, for example, the Agilent 1260 Infinity II SFC System may be used with minimal modification to install a reactor instead of a separation column. Preferably, the solvent or solvent mixtures are selected from: mixtures of organic and aqueous solvents (including halogenated solvents, alcohols, esters, ethers, amines, amides, heterocyclic solvents). It will be appreciated that the solvent mixtures may include any number of solvents as needed, and are not limited only to binary or tertiary mixtures.

[0079] Monitoring one or more process parameters at different solvent ratios

[0080] Any suitable means for monitoring the one or more process parameters may be selected depending on the chemical process.

[0081] By monitoring one or more process parameters, e.g. using analytical tools such as on-line HPLC that measure the level of the input and / or output substances, this information can advantageously be used to determine other performance parameters related to the chemical process(es), e.g. product yield in a chemical reaction. It will be appreciated that combinations of analytical tools may be used to monitor multiple process parameters for any given chemical process or reaction.

[0082] Preferably, the method is quantitative and the one or more process parameters are quantitative parameters.

[0083] If the chemical process is a chemical reaction, the one or more process parameters may be selected from: product yield, reaction species concentration(s), heat of reaction, weight average molecular weight, number average molecular weight, gas offtake, or gas uptake, purity, enantiomeric excess, dispersity, selectivity, rate of reaction, reaction orders, activation energy, or combinations thereof, preferably product yield and / or reaction species concentration(s), most preferably product yield.

[0084] The monitoring of the one or more process parameters may involve monitoring by normal phase, chiral or size exclusion HPLC, and / or detection with other detectors beyond DAD such as a refractive index detector (RID), mass spectrometry detector (MSD), or evaporative light scattering detector (ELSD), or combinations thereof.

[0085] For HPLC analysis or analysis by other analytical tools, sampling can also be performed readily on-line using an autosampler and then the analysis need not be on-line and could be at-line, off-line, or include any form of work up step, purification, or solvent exchange before analysis. The one or more process parameters may be monitored by using one or more spectroscopic or spectrometric tools, either on its own or in combination with chemometric analysis. It will be appreciated that some spectroscopic tools either in-line or on-line may be harder to implement due to the changing background of the spectroscopy due to the changing ratio of solvents. However, this is readily solved by use of chemometric analysis by which either trivial background subtraction could be performed from a blank run (i.e. a solvent ramp with no reactants in) or a more complex approach using the full spectrum and either a PCA (Principal Component Analysis) or more complex ANN (Artificial Neural Network) approach.

[0086] The one or more process parameters may be monitored by using tools including but not limited to in-line / on-line HPLC, flow calorimetry or Ion Mobility Spectrometry-Mass Spectrometry (IMS-MS), Ultraviolet / Visible Spectroscopy (UV / Vis), Fourier Transform Infrared Spectroscopy (FTIR), Near Infrared Spectroscopy (NearIR), Raman Spectroscopy, X-ray Spectroscopy, Fluorescence Spectroscopy, or Nuclear Magnetic Resonance Spectroscopy (NMR). NMR, X- ray, and Raman, in particular, may be used without chemometric approaches as organic solvents are not interfering to the ranges that would be relevant for reaction analysis.

[0087] If the chemical process(es) includes a polymerisation reaction, then for polymer analysis on-line GPC may be used. Alternatively, or in addition, other tools such as flow calorimetry or Ion Mobility Spectrometry-Mass Spectrometry (IMS-MS) may be used.

[0088] Depending on the analytical tool chosen a range of chemical process outcomes could be monitored. For instance, for a chemical reaction, preferably yield and reaction species concentration measurements (e.g. including starting materials, products, and / or by products) are monitored in the method.

[0089] Other chemical process outcomes that may be monitored may depend on the analytical tool chosen and include, but are not limited to, one or more of: purity, enantiomeric excess, dispersity, selectivity, heat of reaction, rate of reaction, gas offtake or uptake, or any combinations thereof.

[0090] It will be appreciated that, if it can be analysed by an analytical tool known in the art and implemented in flow for in situ analysis, or has been implemented at all for ex situ analysis then such a tool may be used with the method herein either through an in-line or on-line analysis approach or be on-line sampling for at-line or off-line analysis. It is most preferred that one or more process parameters are monitored by an on-line chromatographic tool, specifically on-line High-Performance Liquid Chromatography (HPLC) with a Diode-Array Detector (DAD) and Mass Spectrometer (MS) detector is especially preferred. Use of this analytical tool is particularly advantageous for reactions relevant to the pharmaceutical and agrochemical industry, since it allows for quantitative analysis of reaction mixtures of many organic chemical reactions.

[0091] Data sets

[0092] The output from the methods herein may be compiled to provide a data set comprising the one or more process parameters at each solvent ratio. The data set obtained is a continuous numerical data series. Preferably, the data set contains at least reaction time, initial concentrations, temperature, and solvent ratio for each data point.

[0093] The data set obtained allows for the optimal solvent system to be selected to optimise the outcome of the chemical process(es). Optimal in this sense may vary depending on market conditions and regulations on different solvents, changes to other downstream or upstream chemical processes, and environmental incentives. As such the large quantity of diverse solvent reaction data which may be collected by this method may inform optimal solvent choice for such requirements as purity of product, yield of product, environmental factor (e-factor), cost benefit analysis, solvent security, or combinations thereof.

[0094] These data sets are particularly suitable for use in Machine Learning algorithms, with the aim of assisting predictive chemistry. Potentially applications include, but are not limited to kinetic modelling, reaction space exploration, self-optimisation. These methods could use the data to understand and predict solvent effects utilising the efficient data generation of this method to generate large quantities of solvent system data. The data obtained also provides benchmarking for models widely used in the chemical industry such as COSMO-RS.

[0095] Method of training a machine learning model

[0096] In the following discussion, reference is made to Figures 9, 10, 11 , and 12.

[0097] There is provided a method of training a machine learning model to predict one or more process parameters of a chemical process to convert one or more input substances to one or more output substances at different solvent ratios of at least two solvents or for a solvent, the method comprising: monitoring the one or more process parameters at different solvent ratios of one or more pairs of solvents using the method according to the first aspect (101); recording training data pairs of respective solvent ratio data points and associated process parameter data points, the solvent ratio data points comprising a value of solvent ratio and the process parameter data points comprising respective one or more values of the one or more process parameters (102); and training a machine learning model using the training data pairs to output a predicted process parameter data point in response to an input of a solvent ratio data point (103).

[0098] The above method of training a machine learning model can be used to predict one or more process parameters of the chemical process for a single solvent. In this instance, the inputted solvent ratio data point comprises a value of solvent ratio that is 1 (i.e. 100% pure solvent, rather than a solvent mixture). This is advantageous because it allows the user to, for example, recommend an alternative solvent (such as a green alternative, e.g. cyrene or similar) which was not in the solvent training data.

[0099] There is also provided a method of training a machine learning model to predict one or more process parameters of a chemical process to convert one or more input substances to one or more output substances at different solvent ratios of at least two solvents, the method comprising: receiving training data pairs of respective training data points and associated process data parameter points, wherein the training data pairs have been generated by monitoring the one or more process parameters at different solvent ratios of one or more pairs of solvents using the method of any preceding claim, and recording training data pairs of respective solvent ratio data points and associated process parameter data points, the solvent ratio data points comprising a value of solvent ratio and the process parameter data points comprising respective one or more values of the one or more process parameters; and training a machine learning model using the training data pairs to output a predicted process parameter data point in response to an input of a solvent ratio data point.

[0100] As above, the solvent ratio data point, in the situation of outputting a predicted process parameter for a single solvent, may comprise a value of solvent ratio that is 1 (j.e. 100% pure solvent, rather than a solvent mixture).

[0101] The term “value” as used herein can be a scalar or a vector or a tensor, depending on the number of dimensions (e.g. one or more process parameters, one or more ratios, one or more solvent properties, etc.). It will be appreciated that the method of training a machine learning model can be applied for a chemical process involving any number of solvents as needed, and is not limited only to binary or tertiary mixtures of solvents. Accordingly, the solvent ratio data points may comprise a plurality of solvent ratios (e.g. one for each pair of solvents).

[0102] The machine learning model may output a measure of uncertainty for each predicted process parameter data point.

[0103] The machine learning model may be any kernel method, and / or method that provides covariance estimates between input points. Preferably, the machine learning model is a Gaussian Process. The “measure of uncertainty” in the Gaussian Process (GP) model is the covariance matrix for each data point. The GP model outputs a mean and (co)variance for each input data point. To obtain a co-variance matrix this involves predicting multiple data points. The co-variance matrix corresponds to the co-variances between all input data points. The measure of uncertainty may be a variance or a covariance matrix, or may be derived from a variance or a covariance matrix.

[0104] The method may further comprise the steps of: generating a respective first data set for each of at least two further pairs of solvents, each first data set comprising, for the respective pair of solvents, predicted process parameter data points output by the trained machine learning model for respective solvent data points and a respective measure of uncertainty output by the trained machine learning model associated with each predicted process parameter (201 , 301); selecting at least one of the at least two further pairs of solvents based on the measures of uncertainty in the first data sets (202, 303); for each of the selected at least one further pair of solvents, monitoring the one or more process parameters at different solvent ratios of a pair of solvents using the method according to the first aspect (203, 304) and recording further training data pairs of respective solvent ratio data points and associated process parameter data points (204, 305); and training the machine learning model using the further training data pairs to output a predicted process parameter data point in response to an input of a solvent ratio data point (205, 306).

[0105] The step of selecting at least one of the at least two further pairs of solvents based on the measure of uncertainty may comprise selecting at least one of the at least two further pairs of solvents based on a measure of mutual information for each first data set (303). This step may involve comparing a measure of mutual information for each of the first data sets ( / .e. the predicted data sets for untested solvents), and based on this information, selecting at least one of the at least two further pairs of solvents for the monitoring, recording, and training steps, in order to add this information to the machine learning model. This step is advantageous since the at least two further pairs of solvents can be specifically selected to be from a pre-defined space of solvents that are of interest to a user. In other words, this allows the user to choose an area of the solvent space to target beforehand. By selecting at least one of the at least two further pairs of solvents that leads to the largest increase in mutual information, this improves the machine learning model with a minimum number of experiments.

[0106] The method may further comprise predicting a respective second data set for each of the one or more pairs of solvents, each second data set comprising, for the respective pair of solvents, predicted process parameter data points for respective solvent data points and a respective measure of uncertainty associated with each predicted process parameter, wherein the measure of mutual information is a respective gain in mutual information when each first data set is added to the second data set (302). Preferably, the pair of solvents that has the largest measure of mutual information associated with it is selected (303).

[0107] The method may further comprise (i) determining if a stopping criterion is met; and (ii) if the stopping criterion is not met, repeating the above-mentioned steps of generating, selecting, monitoring, recording and training.

[0108] The stopping criterion may comprise a measure of an error in the prediction by the machine learning model being less than or equal to a threshold value. The stopping criterion may comprise the measure of mutual information being less than or equal to a threshold value. This allows for the method of training to be performed until the machine learning model reaches a desired level of accuracy.

[0109] There is also provided a method of designing a chemical process to convert one or more input substances to one or more output substances, the method comprising: using a machine learning model trained according to the method described hereinabove to predict respective one or process parameters for a plurality of ratios of solvents of a pair of solvents (401); selecting a solvent ratio based on the predicted respective one or more process parameters (402); and selecting the solvents of the pair of solvents at the selected solvent ratio for the process

[0110] (403).

[0111] The method may also comprise: using the machine learning model to predict respective one or more process parameters for a plurality of ratios of solvents of a pair of solvents for each of a plurality of pairs of solvents (401); selecting a solvent ratio and pair of solvents based on the predicted respective one or more process parameters (402); selecting the solvents of the selected pair of solvents at the selected solvent ratio for the process (403).

[0112] There is also provided a method for performing a chemical process to convert one or more input substances to one or more output substances (404), the method comprising: providing a mixture comprising the one or more input substances for the chemical process and a solvent mixture, wherein the solvent mixture comprises at least two solvents at a solvent ratio, wherein the pair of solvents and the solvent ratio of the pair of solvents of the chemical process has been selected according to the method described hereinabove; and performing the chemical process thereby converting the one or more input substances to provide the one or more output substances (404).

[0113] There are a number of advantages with the above-mentioned methods, which include (but are not limited to): i) even with a limited data-set the above-mentioned methods provide informationseeking data collection policies which can have a much larger impact when choosing which experiments to run in a larger search space; ii) allowing for more efficient data collection for a chemical process, requiring fewer experiments, and, in turn, also leading to reduced cost and waste; iii) allowing for more complete data collection ( / .e. even with the same number of experiments, the data collected is spread-out in a more useful manner to provide more information rich data for prediction), therefore providing more useful data for decision making during chemical process design; and / or iv) allowing for the selection of solvent alternatives which were not tested for in the initial data collection (training data) and use these alternatives with confidence despite having not physically screened them, due to the information density on the effects of solvent properties on the process outcome obtained. System

[0114] There is provided a system for monitoring one or more process parameters of a chemical process to convert one or more input substances to one or more output substances at different solvent ratios, comprising:

[0115] (i) a plurality of variable flow pumps, each in fluid communication with an input stream comprising a solvent or solvent mixture and the one or more input substances for the chemical process, wherein at least two of the input streams comprise a distinct solvent or solvent mixture;

[0116] (ii) a flow reactor in fluid communication with each of the variable flow pumps; and

[0117] (iii) an analytical unit in fluid communication with the flow reactor, configured to monitor one or more process parameters.

[0118] It will be appreciated that the various features described above in relation to the method of the first aspect apply equally to the system.

[0119] The system preferably comprises a control system configured to control the flow rate of the input streams to the flow reactor via the variable flow pumps.

[0120] The system may be for concurrently monitoring one or more process parameters of multiple chemical processes, wherein the input substances for each chemical process are provided in the plurality of input streams, and the analytical unit is configured to monitor one or more process parameters for each chemical process at different solvent ratios, wherein each input stream comprises the input substances for all of the chemical processes, or each input stream comprises the input substances for one chemical process.

[0121] The analytical unit may include, but is not limited to, in-line / on-line HPLC, flow calorimetry or Ion Mobility Spectrometry-Mass Spectrometry (IMS-MS), Ultraviolet / Visible Spectroscopy (UV / Vis), Fourier Transform Infrared Spectroscopy (FTIR), Near Infrared Spectroscopy (NearIR), Raman Spectroscopy, X-ray Spectroscopy, Fluorescence Spectroscopy, or Nuclear Magnetic Resonance Spectroscopy (NMR). NMR, X-ray, and Raman, in particular, may be used without chemometric approaches as organic solvents are not interfering to the ranges that would be relevant for reaction analysis.

[0122] Preferably, wherein the chemical process is a biphasic liquid-gas process, the flow reactor is a gas-liquid contactor, preferably a tube-in-tube reactor, bubble type reactor, or segmented flow system. The reactor is preferably a tubular reactor or Continuous Stirred-Tank Reactor (CSTR) for homogeneous reactions initiated by heating, cooling, microwave-activated, photoactivated, or simply initiated through mixing (or a combination of the same). For a heterogeneous process or reaction (involving liquid-solid, or liquid-solid-gas), the reactor is preferably a packed bed reactor or Continuous Stirred-Tank Reactor (CSTR). Most preferably, the flow reactor is a thermally heated tubular reactor with fluid dynamics, e.g. within the assumptions of an ideal plug flow reactor.

[0123] If the chemical process(es) includes a polymerisation reaction, then for polymer analysis on-line GPC may be used. Alternatively, or in addition, other tools such as flow calorimetry or IMS-MS may be used.

[0124] Preferably, the analytical unit may be High-performance Liquid Chromatography (HPLC), more preferably selected from normal phase, reverse phase, chiral, or size exclusion HPLC. The HPLC may comprise one or more of a diode array detector (DAD), refractive index detector (RID), mass spectrometry detector (MSD), or evaporative light scattering detector (ELSD).

[0125] The aspects provided herein are also described in the following clauses:

[0126] 1. A method for monitoring of one or more process parameters of a chemical process to convert one or more input substances to one or more output substances at different solvent ratios, comprising: a) providing a plurality of input streams, each comprising a solvent or solvent mixture and one or more input substances for the chemical process, wherein at least two of the input streams comprise a distinct solvent or solvent mixture; b) delivering one or more of the input streams to a flow reactor; c) varying the ratio of solvents within the flow reactor by changing relative flow rates of the input streams to the flow reactor; and d) monitoring one or more process parameters at different solvent ratios.

[0127] 2. A method according to clause 1 , wherein varying the ratio of solvents comprises a continuous linear or non-linear ramping of the relative flow rates.

[0128] 3. A method according to clause 1 or clause 2, wherein the chemical process is selected from a chemical reaction, an extraction, a crystallisation, and a quenching step.

[0129] 4. A method according to any one of clauses 1 to 3, wherein the method is for concurrently monitoring one or more process parameters of multiple chemical processes, wherein for step (a) the input substances for each chemical process are provided in the plurality of input streams, and step (d) comprises monitoring one or more process parameters for each chemical process at different solvent ratios, wherein each input stream comprises the input substances for all of the chemical processes, or each input stream comprises the input substances for one chemical process.

[0130] 5. A method according to any one of the preceding clauses, wherein the method is quantitative and the one or more process parameters are quantitative parameters.

[0131] 6. A method according to any one of the preceding clauses, wherein in step (c) the ratio of solvents is varied whilst maintaining a constant total flow rate to the flow reactor.

[0132] 7. A method according to any one of the preceding clauses, wherein the chemical process is a chemical reaction to produce one or more products from one or more starting materials, wherein the one or more input substances are one or more starting materials and the one or more output substances are the one or more products. 8. A method according to clause 7, wherein the method comprises (e) determining reaction performance from the monitored one or more process parameters at different solvent ratios.

[0133] 9. A method according to clause 7, wherein the one or more process parameters are selected from: product yield, reaction species concentration, heat of reaction, weight average molecular weight, number average molecular weight, gas offtake, or gas uptake, purity, enantiomeric excess, dispersity, selectivity, rate of reaction, reaction orders, activation energy, or combinations thereof.

[0134] 10. A method according to any one of the preceding clauses, comprising three or more input streams, wherein three of the input streams comprise a distinct solvent or solvent mixture.

[0135] 11. A method according to any one of the preceding clauses, wherein the chemical process is a unimolecular, bimolecular or multicomponent chemical reaction.

[0136] 12. A method according to any one of the preceding clauses, wherein the chemical process is a chemical reaction selected from a homogeneous organic chemical reaction, a heterogenous catalysis reaction, a homogeneous catalysis reaction, a biphasic liquid-gas reaction, a nano particle synthesis reaction, a crystallisation reaction, a biological reaction, a fermentation reaction, or a polymerisation reaction.

[0137] 13. A method according to clause 11 , wherein the chemical process is a homogeneous organic chemical reaction between multiple reactants with or without a catalyst.

[0138] 14. A method according to clause 13, wherein the method is for concurrently monitoring one or more process parameters of multiple homogeneous organic chemical reactions between multiple reactants with or without a catalyst, wherein for step (a) the input substances for each chemical reaction are provided in the plurality of input streams, step (d) comprises monitoring one or more process parameters for each chemical reaction at different solvent ratios, and wherein each input stream comprises the input substances for all of the chemical reactions, or each input stream comprises the input substances for one chemical reaction. 15. A method according to any one of the preceding clauses, wherein the solvent or solvent mixtures comprise a liquid at room temperature and pressure.

[0139] 16. A method according to any one of the preceding clauses, wherein the chemical process is a biphasic liquid gas reaction, and the flow reactor is a gas-liquid contactor, preferably a tube-in-tube reactor, bubble type reactor, or segmented flow system.

[0140] 17. A method according to any one of the preceding clauses, wherein the solvent or solvent mixtures of the plurality of input streams comprise a supercritical fluid.

[0141] 18. A method according to any one of the preceding clauses, wherein the solvent or solvent mixtures of the plurality of input streams comprise a deep eutectic solvent and / or inorganic derived solvent.

[0142] 19. A method according to any one of the preceding clauses, wherein the solvent or solvent mixtures of the plurality of input streams comprise an organic solvent selected from one or more of polar, apolar, protic, aprotic, halogenated, aqueous, and aromatic solvents.

[0143] 20. A method according to any preceding clause, wherein the solvent or solvent mixtures are selected from mixtures of organic and aqueous solvents.

[0144] 21. A method according to any one of the preceding clauses, wherein the one or more process parameters are monitored by in-line or on-line monitoring.

[0145] 22. A method according to any one of clauses 1 to 20, wherein the one or more process parameters are monitored by at-line or off-line monitoring.

[0146] 23. A method according to any one of the preceding clauses, wherein the one or more process parameters are monitored by using High-performance Liquid Chromatography (HPLC).

[0147] 24. A method according to clause 23, wherein the one or more process parameters are measured by on-line HPLC with a DAD and MS detector.

[0148] 25. A method according to clause 23, wherein the one or more process parameters are monitored using normal phase, reverse phase, chiral, or size exclusion HPLC. 26. A method according to clause 23 or clause 25, wherein the HPLC comprises one or more of a diode array detector (DAD), refractive index detector (RID), mass spectrometry detector (MSD), or evaporative light scattering detector (ELSD).

[0149] 27. A method according to any one of clauses 1 to 22, wherein the one or more process parameters are monitored by using one or more spectroscopic or spectrometric tools, either on its own or in combination with chemometric analysis.

[0150] 28. A method according to clause 27, wherein the tool is selected from the list of UV / Vis, FTIR, NearIR, Raman, X-ray, fluorescence, NMR, or Mass Spectrometry.

[0151] 29. A method according to any one of clauses 1 to 22, wherein the one or more process parameters are monitored by using thermography, flow calorimetry, or gas chromatography.

[0152] 30. A method according to any one of the clauses 1 to 11 , wherein: i) the chemical process is a homogeneous organic chemical reactions between multiple reactants with or without a catalyst; ii) the solvent or solvent mixtures are selected from mixtures of organic and aqueous solvents; and iii) the one or more process parameters are measured by an on-line chromatographic tool, preferably on-line HPLC with a DAD and MS detector.

[0153] 30a. A method according to any one of the clauses 1 to 11 , wherein: i) the chemical process is a heterogeneous organic chemical reactions between multiple reactants with or without a catalyst; ii) the reactor is a packed bed reactor system; iii) the solvent or solvent mixtures are selected from mixtures of organic and aqueous solvents; and iv) the one or more process parameters are measured by an on-line chromatographic tool, preferably on-line HPLC with a DAD and MS detector.

[0154] 30b. A method according to any one of the clauses 1 to 6 and 10, wherein: i) the chemical process is an extraction of a compound from a mixture of compounds, either as a reaction quenching step or extraction of value added chemicals from biomass or recycled waste; ii) the solvent or solvent mixtures are selected from mixtures of organic and aqueous solvents; and iii) the one or more process parameters are measured by an on-line chromatographic tool, preferably on-line HPLC with a DAD and MS detector.

[0155] 31. A method according to any of clauses 30, 30a, or 30b, wherein the method is for concurrently monitoring one or more process parameters of multiple homogeneous organic chemical reactions between multiple reactants with or without a catalyst, wherein for step (a) the input substances for each chemical reaction are provided in the plurality of input streams, step (d) comprises monitoring one or more process parameters for each chemical reaction at different solvent ratios, and wherein each input stream comprises the input substances for all of the chemical reactions, or each input stream comprises the input substances for one chemical reaction.

[0156] 32. A method according to clause 30, 30a, 30b, or 31 , wherein the one or more process parameters are yield and / or reaction species concentration measurements.

[0157] 33. A method according to any one of the preceding clauses, comprising compiling a data set comprising the one or more process parameters at each solvent ratio for the one or more chemical processes.

[0158] 34. A method according to clause 33, wherein the data set is provided to a Machine Learning (ML) algorithm.

[0159] 35. A method of training a machine learning model to predict one or more process parameters of a chemical process to convert one or more input substances to one or more output substances at different solvent ratios of at least two solvents or for a solvent, the method comprising: monitoring the one or more process parameters at different solvent ratios of one or more pairs of solvents using the method of any preceding claim; recording training data pairs of respective solvent ratio data points and associated process parameter data points, the solvent ratio data points comprising a value of solvent ratio and the process parameter data points comprising respective one or more values of the one or more process parameters; and training a machine learning model using the training data pairs to output a predicted process parameter data point in response to an input of a solvent ratio data point. 35a. A method of training a machine learning model to predict one or more process parameters of a chemical process to convert one or more input substances to one or more output substances at different solvent ratios of at least two solvents or a solvent, the method comprising: receiving training data pairs of respective training data points and associated process data parameter points, wherein the training data pairs have been generated by monitoring the one or more process parameters at different solvent ratios of one or more pairs of solvents using the method of any preceding claim, and recording training data pairs of respective solvent ratio data points and associated process parameter data points, the solvent ratio data points comprising a value of solvent ratio and the process parameter data points comprising respective one or more values of the one or more process parameters; and training a machine learning model using the training data pairs to output a predicted process parameter data point in response to an input of a solvent ratio data point.

[0160] 36. The method of clause 35 or 35a, wherein the machine learning model outputs a measure of uncertainty for each predicted process data point.

[0161] 37. The method of clause 36, wherein the machine learning model is a kernel method, and / or method that provides co-variance estimates between input points, or preferably is a Gaussian Process.

[0162] 38. The method of clause 36 or 37, wherein the measure of uncertainty is or is derived from a variance or a covariance matrix.

[0163] 39. The method of clause 36, 37 or 38, comprising: generating a respective first data set for each of at least two further pairs of solvents, each first data set comprising, for the respective pair of solvents, predicted process parameter data points output by the trained machine learning model for respective solvent data points and a respective measure of uncertainty output by the trained machine learning model associated with each predicted process parameter; selecting at least one of the at least two further pairs of solvents based on the measures of uncertainty in the first data sets; for each of the selected at least one further pair of solvents, monitoring the one or more process parameters at different solvent ratios of a pair of solvents using the method of any one of clauses 1 to 24 and recording further training data pairs of respective solvent ratio data points and associated process parameter data points; and training the machine learning model using the further training data pairs to output a predicted process parameter data point in response to an input of a solvent ratio data point.

[0164] 40. The method of clause 39, wherein selecting at least one of the at least two further pairs of solvents based on the measure of uncertainty comprises selecting at least one of the at least two further pairs of solvents based on a measure of mutual information for each first data set.

[0165] 40a. The method of clause 40, where in the measure of mutual information is local mutual information.

[0166] 41. The method of clause 40 or 40a, comprising predicting a respective second data set for each of the one or more pairs of solvents, each second data set comprising, for the respective pair of solvents, predicted process parameter data points for respective solvent data points and a respective measure of uncertainty associated with each predicted process parameter, wherein the measure of mutual information is a respective gain in mutual information when each first data set is added to the second data set.

[0167] 42. The method of clause 41 , wherein the pair of solvents that has the largest measure of mutual information associated with it is selected.

[0168] 43. The method of any one of clauses 39 to 42: determining if a stopping criterion is met; and if the stopping criterion is not met, repeating the generating, selecting, monitoring, recording and training.

[0169] 44. The method of clause 43, wherein the stopping criterion comprises a measure of an error in the prediction by the machine learning model being less than or equal to a threshold value.

[0170] 45. The method of clause 43, when dependent on clause 40, 41 or 42, wherein the stopping criterion comprises the measure of mutual information being less than or equal to a threshold value.

[0171] 45a. The method of clause 43, when dependent on clause 40, 41 , or 42, wherein the stopping criterion comprises a maximum predictive variance being less than or equal to a threshold value. 46. A method of designing a chemical process to convert one or more input substances to one or more output substances, the method comprising using a machine learning model trained according to the method of any one of clauses 35 or 35a to 45 to predict respective one or process parameters for a plurality of ratios of solvents of a pair of solvents; selecting a solvent ratio based on the predicted respective one or more process parameters; selecting the solvents of the pair of solvents at the selected solvent ratio for the process.

[0172] 47. The method of clause 46 comprising using the machine learning model to predict respective one or more process parameters for a plurality of ratios of solvents of a pair of solvents for each of a plurality of pairs of solvents; selecting a solvent ratio and pair of solvents based on the predicted respective one or more process parameters; selecting the solvents of the selected pair of solvents at the selected solvent ratio for the process.

[0173] 48. A method for performing a chemical process to convert one or more input substances to one or more output substances, the method comprising: providing a mixture comprising the one or more input substances for the chemical process and a solvent mixture, wherein the solvent mixture comprises at least two solvents at a solvent ratio, wherein the pair of solvents and the solvent ratio of the pair of solvents of the chemical process has been selected according to the method of claim 46 or 47; and performing the chemical process thereby converting the one or more input substances to provide the one or more output substances.

[0174] 49. A system for monitoring one or more process parameters of a chemical process to convert one or more input substances to one or more output substances at different solvent ratios, comprising:

[0175] (i) a plurality of variable flow pumps, each in fluid communication with an input stream comprising a solvent or solvent mixture and the one or more input substances for the chemical process, wherein at least two of the input streams comprise a distinct solvent or solvent mixture;

[0176] (ii) a flow reactor in fluid communication with each of the variable flow pumps; and (iii) an analytical unit in fluid communication with the flow reactor, configured to monitor one or more process parameters.

[0177] 50. The system according to clause 49, comprising a control system configured to control the flow rate of the input streams to the flow reactor via the variable flow pumps.

[0178] 54. The system according to clause 49 or 50, wherein the system is for concurrently monitoring one or more process parameters of multiple chemical processes, wherein the input substances for each chemical process are provided in the plurality of input streams, and the analytical unit is configured to monitor one or more process parameters for each chemical process at different solvent ratios, wherein each input stream comprises the input substances for all of the chemical processes, or each input stream comprises the input substances for one chemical process.

[0179] 52. The system according to any one of clauses 4948 to 51 , wherein the chemical process is a homogeneous reaction initiated by one or more of heating, cooling, microwave-activated, photo-activated, or simply initiated through mixing, and the flow reactor is selected from a tubular reactor or Continuous Stirred-Tank Reactor (CSTR).

[0180] 53. The system according to any one of clauses 49 to 51 , wherein the chemical process is a biphasic liquid-gas process and the flow reactor is a gas-liquid contactor, preferably a tube-in-tube reactor, bubble type reactor, or segmented flow system.

[0181] 54. The system according to any one of clauses 49 to 51 , wherein the chemical process is a heterogeneous process, the flow reactor is preferably a packed bed reactor or Continuous Stirred-Tank Reactor (CSTR).

[0182] 55. The system according to any one of clauses 49 to 52, wherein the flow reactor is a thermally heated tubular reactor.

[0183] 56. The system according to any one of clauses 49 to 55, wherein the analytical unit comprises one or more spectroscopic or spectrometric tools, either on its own or in combination with chemometric analysis, preferably selected from in-line / on-line Ultraviolet / Visible Spectroscopy (UV / Vis), Fourier Transform Infrared Spectroscopy (FTIR), Near Infrared Spectroscopy (NearIR), Raman Spectroscopy, X-ray Spectroscopy, Fluorescence Spectroscopy, or Nuclear Magnetic Resonance Spectroscopy (NMR). 57. The system according to any one of clauses 49 to 56, wherein the analytical unit comprises High-performance Liquid Chromatography (HPLC), preferably selected from normal phase, reverse phase, chiral, or size exclusion HPLC.

[0184] 58. The system according to clause 57, wherein the HPLC comprises one or more of a diode array detector (DAD), refractive index detector (RID), mass spectrometry detector (MSD), or evaporative light scattering detector (ELSD).

[0185] 59. The system according to any one of clauses 49 to 58, wherein the chemical process is optionally a polymerisation reaction, and the analytical unit comprises thermography, flow calorimetry, or gas chromatography, or a combination thereof.

[0186] Examples

[0187] The present invention will now be described by way of reference to the following examples.

[0188] These examples are not to be construed as being limiting on the invention.

[0189] Example 1 : Unimolecular chemical reaction

[0190] 1 .1 Synthesis of compounds

[0191] 1 and 3 were purchased from commercial sources (Sigma-Aldrich) and used as received.

[0192] 2 was synthesised from 1 for calibration purposes. 200 pL of 1 was dissolved in 9 mL ethanol and 5 mL of water using a pipette (0.283 g). The flow system was primed with the reaction solution for 2 min at 1 mL min-1 and the reactor was heated to 220 °C. The reaction solution was pumped at 0.2 mL min-1 (16 min residence time) at 220 °C at 70 bar for 23 min. The reactor effluent was then collected for 70 min with addition of ethanol to the inlet at 35 min. The collected solution was evaporated to half volume in vacuo to remove the ethanol. The resultant solution was dissolved in DCM (10 mL) and 2.5 M NaOH (10 mL) was added. The separated aqueous solution was acidified with 2 M HCI (15 - 25 mL) and washed with DCM (10 mL). The DCM was evaporated in vacuo to afford 2.

[0193] 1 .2 Flow system overview

[0194] An example of a flow system is shown in Figure 1 , and was constructed in which reagent(s) and solvent(s) as solutions are delivered from Duran bottles fitted with inlet tubing with in-line filters. The inlet tubing lines (1 / 8” OD, 2 mm ID PTFE, 1.6 mL and 1.2 mL respectively) were connected to a Gilson 305 HPLC pump (pump A) and a Gilson 307 HPLC pump (pump B) respectively. The outlets of these pumps were connected to two stainless steel tubing (25 cm) which fed into a Valeo T-piece stainless steel mixer.

[0195] The outlet of the mixer is connected to a 5.14 m length of stainless steel tubing (1 mm ID, 1 / 16” OD); the first 30 cm of which are outside of the GC oven (heated reactor length: 4.84 m, heated reactor volume: 3.92 mL). This is then connected by a stainless steel HPLC style union at the outlet of the GC oven through an insulating PTFE sheet to a length of stainless steel tubing (14 cm) which passes through a custom built cooling system consisting of an aluminium block and a Peltier assembly. The outlet of the cooling device is connected by a stainless steel HPLC style union to a 20 cm length of stainless steel tubing into a VALCO variable pressure Back Pressure Regulator (BPR). The BPR outlet is connected to a 25 cm length of stainless steel tubing which connects to a VALCO T-piece stainless steel mixer. The other mixer fittings are connected to the inlet to the modified HPLC 6-port, 2-way injector valve and to the dilution pump of the HPLC setup.

[0196] This flow system is a modification of a system previously used and which was previously thoroughly assessed for residence time distributions, as mentioned in the publication L. Schrecker, J. Dickhaut, C. Holtze, P. Staehle, M. Vranceanu, K. Hellgardt, K. K. Hii, React. Chem. Eng. 2022, 8, 41^6.

[0197] 1.3 Pumps

[0198] The pumps used are a Gilson 305 HPLC pump and a Gilson 307 HPLC pump with a 10 mL WSC and 10 mL SC pump heads respectively. The inlet and outlet fittings to the Gilson pumps are standard Gilson 30X fittings. All other fittings are HPLC style fittings. These are connected by a GSIOC cable to allow pump control through the inbuilt Gilson firmware. The pressure of the system is monitored by the manometric module in the Gilson 307 pump. In order to maintain the pumps in working order and confirm their accuracy, the check valves are cleaned and sonicated in methanol regularly, and the cumulative flow rate of the system is confirmed at different flow rates and at different times during experimentation.

[0199] 1.4 Heating and cooling

[0200] Heating is performed by an HP 5890 Series II GC oven, using the inbuilt software / PID to allow accurate control over temperature ramping. Cooling is performed by a 60 W Peltier thermoelectric cooler module and heatsink assembly (PiHut) with a custom cooling block milled to fit 1 / 16” tubing by the Advanced HackSpace, Imperial College London.

[0201] 1.5 On-line HPLC

[0202] On-line analysis was implemented by modifying the injector on an Agilent HP 1100 series HPLC. The outlet stream of the reactor is diluted (when required) by one of the Agilent HP 1100 series binary HPLC pumps at a Valeo T-piece stainless steel mixer. The other fitting goes into the 6-port, 2-way injector valve. The fittings to the 6-port, 2-way injector valve were disconnected and reconnected as shown in the scheme above to allow a 1.8 pL sample loop to be continually filled from the reactor effluent in the resting valve position. When triggered by the Chemstation HPLC software the injector method switches the valve to its second position for 0.1 min, allowing the contents of the sample loop to be redirected by solvent from the Agilent HP 1100 series pumps into the column and subsequent detector.

[0203] The HPLC column used was a Restek Raptor C182.7pm 50 mm x 2.1 mm running an isocratic mixture of 40% acetonitrile in water. This method allowed injections every 2 minutes.

[0204] Table 3. Key parameters set for the HPLC method used in this example for on-line HPLC analysis.

[0205] The Agilent 1100 Autosampler 1 automatically switches the valve at the beginning of the injector program. The inputted injector cycle is a “WAIT” command for 0.10 min and then a “VALVE MAIN PASS” command.

[0206] Figure 2 shows an example chromatogram obtained according to the HPLC method above for monitoring a unimolecular chemical reaction.

[0207] Table 4. Substituents for phenyl allyl ether substrates and their corresponding retention times in the HPLC method in Figure 2.

[0208] 1 .6 Calibration method

[0209] The HPLC method was used to determine quantitative concentrations of starting materials and products by generating calibration curves based on known concentration samples. Starting material and product peak areas were measured relative to the area of the internal standard peak (ethyl benzene, 3). Ethyl benzene was chosen as a non-interacting internal standard that is UV visible, thermally stable, stable to solvolysis and separable from the starting materials and products by liquid chromatographical methods.

[0210] Multiple solutions were prepared, using pipettes and volumetric flasks, containing varying quantities of 1 , 2, and 3. Dilutions of these solutions were injected into the HPLC in a manual method to give 5 point calibration curves. Another method was also used for calibration by which a solution containing 1 and / or 2 at a known concentration in ethanol was pumped through the flow system, with simultaneous delivery of a solution of 3 in ethanol by the second pump. By varying the ratio of flow rates of the two pumps in a linear manner, whilst maintaining the same cumulative flow rate, it was possible to rapidly generate calibration curves with greater than 10 calibration points each using the on-line HPLC setup.

[0211] Figure 3 shows an example HPLC calibration curve collected via a flow rate ramp between a solution containing starting materials including 1d (Z = H) and a solution containing ethyl benzene as an internal standard.

[0212] 1x

[0213] 1.7 Solvent ramping method

[0214] Multiple solutions were prepared containing 1 and 3 in known and equal quantities dissolved in different solvents or solvent mixtures. These were prepared either as 50 mL or 100 mL solution in volumetric flasks. Two of these solutions were attached to the inlet lines of the flow system and both pumps were primed for at least 3 minutes at 2 mL min-1. The flow rate was dropped to 0.2 mL min-1and the reactor temperature raised to 220 °C at 100 bar. A ramp was then controlled between the flow rate of the two pumps by the Gilson firmware either as a 25 minute ramp or 50 min ramp as described in tables 5A and 5B below. Table 5A

[0215] Table 5B

[0216] Solvent ramps were performed in this manner between a range of different solutions having different solvent or solvent mixtures as described in Table 6 below. The letter in the box corresponds to the subfigure of the results that follow in Figure 4.

[0217] Table 6

[0218] The HPLC samples collected are then automatically analysed using an automated python script which allows each sample to be assigned a time stamp and for the peak areas of 1, 2, and 3 to picked and collated. This data can then be stored as a csv file or visualised as seen in figure 4. The data series represented are for a bidirectional ramp in solvent composition to confirm that the data at each solvent composition is independent of the ramp direction (hysteresis).

[0219] It can be seen that the outputs of the above method are a series of data points for the % yield of the chemical reaction, across a spectrum of the ratios of solvent or solvent mixtures. This method allows, for the first time, for this data to be collected in a continuous manner and in bulk. Figure 4 J.2 demonstrates how side products can also be tracked in changing solvent conditions.

[0220] This provides a number of advantages. The output provides richer information on the performance of the reaction in a variety of different solvent systems. A particular advantage is that this data can then be used to quickly and efficiently select the most appropriate solvent system for the chemical reaction.

[0221] Each data point in these series (in Fig. 4) would previously have required an individual experiment to be performed either by batch or flow methods. Inherently the data collected according to this method is more accurate and reliable than the same data collected in batch. It is more accurate because it prevents solvent evaporation at elevated temperatures by allowing performing of the chemical reaction under pressure and without headspace. There is also more homogenous heating across the reaction solutions due to the large surface area to volume ratio. The method is also safer than carrying out the same steps in batch as it allows for a smaller reactor volume.

[0222] Example 2: Multiple chemical reactions in one pot

[0223] 2.1 Compound synthesis for one-pot example CH3

[0224] 1d and 3 were purchased from commercial sources and used as received, all other compounds were synthesised and purified as listed below. Chemicals described below for the synthesis of other 1x compounds were purchased from Sigma-Adrich, Acros Organic, TCI, and Fluorochem and used as received.

[0225] Synthesis of 4-methoxy phenyl ally ether (1a)

[0226] In a 500 mL RB flask fitted with a reflux condenser, a solution of 4-methoxyphenol (10.1 g, 81.4 mmol) in acetone (250 mL) was stirred at 300 rpm. Potassium carbonate (22.3 g, 161 mmol) was added in batches and the solution was heated to 60 °C under nitrogen and left to stir for 20 minutes. Allyl bromide (9.80 g, 81.0 mmol) was then added and the solution was heated to reflux (66 °C) and left to stir for 4 hours. After cooling to ambient temperature, the solution was filtered through celite and evaporated in vacuo. The residue was purified by automated column chromatography in 10% ethyl acetate in petroleum ether. This produced the desired product as a colourless oil (9.05 g, 44.8 mmol, 55%).

[0227] Synthesis of starting materials in medium throughput (1b, 1c, 1e, 1f)

[0228] The precursors were prepared in parallel using a Reactarray Barnstead RAR-01 OA reaction station equipped with 10 x 25 mL reactors that are fitted individually with reflux condensers. Each reactor was flushed with N2 and charged with a cross-shaped stirrer, potassium carbonate (2.50 g, 18.1 mmol), the requisite phenol starting material (approximately 6 mmol., see table 7 for quantities) and dry acetone (20 mL). The reaction mixtures were subsequently stirred with gentle heating at 30 °C for 20 minutes in order to ensure deprotonation of the phenol. Allyl bromide (1.13 g, 9.34 mmol) was added to each reactor via a pipette, and the reaction mixtures were left stirring at 30 °C for 18 hours. After cooling, the solutions were filtered through celite into falcon tubes and the acetone was removed using a nitrogen blow down evaporator. This produced the desired products as listed below in Table 7.

[0229] Table 7. Values for the medium throughput synthesis of para substituted phenyl allyl ether 1.

[0230] Synthesis of 2-allyl-4-methoxy phenol (2a)

[0231] °\

[0232] 4-Methoxy phenyl allyl ether (1a, 1.60 g, 9.75 mmol) was added to a 0.5 - 2 mL microwave vial with a stirrer. The vial was subjected to microwave heating at 200 °C for 40 min. After cooling to room temperature, the crude product formed was dissolved in petroleum ether (10 mL) and extracted into a 2.5 M NaOH solution (2 x 5 mL). The combined aqueous layers were washed with petroleum ether (2 x 5 mL) before re-acidifying - with cone. HCI (6 mL) to pH 2. The acidified aqueous layer was extracted with ethyl acetate (3 x 10 mL). The combined ethyl acetate washes were diluted with petroleum ether (40 mL) and filtered through a silica plug. The organic mixture was dried over magnesium sulfate, filtered and evaporated in vacuo to produce the desired product as a pale yellow oil (1.33 g, 8.11 mmol, 83%).

[0233] Synthesis of other aromatic Claisen products (2b-h)

[0234] 2d-f:

[0235] 200 pL of each allyl ethers (1d-f) were dissolved in 9 mL ethanol and 5 mL of water using a pipette in quantities listed below in Table 8. The solutions were then pumped through the flow system at 0.2 mL min-1at 220 °C (16 min residence time). Prior to each experiment, the system was primed with the reaction solution for 2 min at 1 mL min-1and the reactor was heated to 220 °C. The reaction solution was pumped at 0.2 mL min-1at 70 bar for 23 min (the time taken for solution to pass from the mixer to the outlet of the system) The reactor effluent was then collected for 70 min with addition of ethanol to the inlet at 35 min. After each experiment the system was then flushed with ethanol. Each solution was evaporated to half volume in vacuo to remove the ethanol. The resultant solution was dissolved in DCM (10 mL) and 2.5 M NaOH (10 mL) was added. The separated aqueous solution was acidified with 2 M HCI (15 - 25 mL) and washed with DCM (10 mL). The DCM was evaporated in vacuo to afford the desired products.

[0236] 2b:

[0237] Attempts to produce 2b (Z = Me) by the above method failed to furnish clean products even after flash column chromatography, as the product either appeared to either remain on the column (silica or alumina) or degrade. Hence, the Claisen rearrangement of 1 b (0.281 g) was performed in flow, by dissolving the precursor in DMF (10 mL) and subjected to heating at 240 °C using the flow system for 4 hours (recirculatory mode, 0.5 mL min-1). Diethyl ether (20 mL) was then added, and the resultant solution was washed with saturated lithium chloride solution (5 x 20 mL). The diethyl ether was then evaporated to give the desired product as a pale yellow oil (0.205 g, 1.38 mmol, 73%)

[0238] 2c: 200 pL of allyl ether 1c was dissolved in ethanol (20 mL) in quantity listed below in Table 8. The reaction solution was subjected to heating at 250 °C using the flow system for 1 hours (recirculatory mode, 0.5 mL min-1). The ethanolic solution is then collected and evaporated in vacuo. The residue is dissolved in DCM (10 mL) and 2.5 M NaOH (10 mL) was added. The separated aqueous solution was acidified with 2 M HCI (25 mL) and washed with DCM (10 mL). The DCM was evaporated in vacuo to afford the desired products.

[0239] Table 8. Values for the medium throughput synthesis of aromatic Claisen products 2.

[0240] 2.2 On-line HPLC Method for one-pot example

[0241] HPLC method

[0242] The column used was a Restek Raptor C182.7pm 50 mm x 2.1 mm running a gradient method in water and acetonitrile as described in Table 9. In the method, the time after the gradient allowed the column to re-equilibrate to a steady state pressure before the following injection. This method allowed injections every 6.5 minutes.

[0243] Table 9. Key parameters set for the HPLC method used in this work for on-line HPLC analysis.

[0244] The Agilent 1100 Autosampler 1 automatically switches the valve at the beginning of the injector program. The inputted injector cycle is a “WAIT” command for 0.10 min and then a “VALVE MAIN PASS” command. Figure 5 shows an example chromatogram of the “one-pot” method using the HPLC method discussed above. The lower diagram shows a diagrammatic representation of the HPLC method.

[0245] Table 10. Substituents for phenyl allyl ether substrates and their corresponding retention times in the HPLC method in Figure S2.

[0246] 2.3 Solvent ramping method for one-pot

[0247] A solution was prepared containing 1 a-f and 3 in known quantities (as listed in T able 11 below) dissolved in absolute ethanol in an 100 mL volumetric flask . This solution was attached to the A pump inlet line of the flow system and a solution of filtered distilled water was attached to the B pump inlet line. Both pumps were primed for at least 3 minutes at 2 mL min-1. The flow rate was dropped to 0.2 mL min-1and the reactor temperature raised to 200 °C at 100 bar. A ramp was then controlled between the flow rate of the two pumps by the Gilson firmware as described in the below Table 12.

[0248] Table 11 : quantities of compounds used Table 12

[0249] The HPLC samples collected were then automatically analysed using an automated python script which allows each sample to be assigned a time stamp and for the peak areas of 1a-f, 2a-f , and 3 to be picked and collated. This data can then be stored as a csv file or visualised as seen in figure 6 as yield of each reaction vs. % water in ethanol.

[0250] The output of the above method is a series of data points for the % yield of all six chemical reactions across a spectrum of the ratios of solvent or solvent mixtures. The method allows, for the first time, for the data relating to multiple reactions to be collected in a continuous manner and in bulk.

[0251] A particular advantage is that this data can then be used to quickly and efficiently select the most appropriate solvent system for each of the six chemical reactions, from performing the above method once.

[0252] Each data point in these series would previously have required an individual experiment to be performed either by batch or flow methods. Data collected according to this method is more accurate and reliable than the same data collected in batch, and allows more homogenous heating across the solutions due to the large surface area to volume ratio. The method is also safer due to a smaller reactor volume.

[0253] Example 3: Bimolecular chemical reaction

[0254] This example demonstrates the applicability of the method of the invention to bimolecular chemical reactions to produce two products from two starting materials, in this example the reaction is a competitive amide bond synthesis with ester formation. 3.1 Materials for amide example

[0255] 3, 4, 5, 6 and 7 were all purchased from commercial sources (Sigma-Aldrich with the exception of 4 which was purchased from Scientific Laboratory Supplies) and used as received for experiments and HPLC calibration.

[0256] 3.2 On-line HPLC for amide example

[0257] The column used was a Restek Raptor C182.7pm 50 mm x 2.1 mm running a gradient method in water and acetonitrile as described in Table S3. In the method, the time after the gradient allowed the column to re-equilibrate to a steady state pressure before the following injection. This method allowed injections every 5.5 minutes. In this example the dilution pump was used to dilute the sample stream by 10:1 with acetonitrile.

[0258] Table 13. Key parameters set for the HPLC method used in this example for on-line HPLC analysis.

[0259] The Agilent 1100 Autosampler 1 automatically switches the valve at the beginning of the injector program. The inputted injector cycle is a “WAIT” command for 0.10 min and then a “VALVE MAIN PASS” command. Figure 7 shows an example chromatogram of the HPLC method discussed above. The lower two diagrams shows a diagrammatic representation of the HPLC pump method and the dilution pump method.

[0260] Table 14. Reaction product and internal standard retention times in the HPLC method in Figure S2.

[0261] 3.3 Solvent ramping method for direct thermal amide synthesis

[0262] Two solutions were prepared containing 3, 4 and 5 in known quantities dissolved in absolute ethanol or 50% water in ethanol in 50 mL volumetric flasks as described in Table 15. The ethanolic solution was attached to the A pump inlet line of the flow system and a the ethanol water mixture solution was attached to the B pump inlet line. Both pumps were primed for at least 3 minutes at 2 mL min-1. The flow rate was dropped to 0.1 mL min-1and the reactor temperature raised to 280 °C at 100 bar. A ramp was then controlled between the flow rate of the two pumps by the Gilson firmware as described in Table 16 below.

[0263] Table 15

[0264] Table 16 The HPLC samples collected were then automatically analysed using an automated python script which allows each sample to be assigned a time stamp and for the peak areas of 6, 7 and 3 to be picked and collated. This data can then be stored as a csv file or visualised as seen in Figure 8 as concentration of each reaction product vs. % water in ethanol.

[0265] It can be seen that there is a suppression of the formation of the amide more than the ester when more water is added, when it would be expected that ester formation would be surpassed more when there is less of the ethanol needed to react.

[0266] The output of the above method is a series of data points for the concentration of the products 6 and 7 of the bimolecular chemical reaction across a spectrum of the ratios of solvent or solvent mixtures. The method allows, for the first time, for this data to be collected in a continuous manner and in bulk.

[0267] This output provides richer information on the performance of the reaction in a variety of different solvent systems. This data can then be used to quickly and efficiently select the most appropriate solvent system for the chemical process.

[0268] Each data point in these series would previously have required an individual experiment to be performed either by batch or flow methods. Inherently the data collected according to the method herein is more accurate and reliable than the same data collected in batch. There is more homogenous heating across the solutions due to the large surface area to volume ratio. The method is safer as it allows for a smaller reactor volume.

[0269] Example 4: Training of a machine learning model

[0270] Introduction

[0271] The “claisen data-set" consists of multiple solvent ramps collected in transient flow. The input space of the problem corresponds to the solvent being used in the reaction. There is a total of 11 base solvents, and new ones are created by combining them: where each Sj corresponds to how much of solvent i is used for the reaction. The focus is on combinations of two solvents and it is denoted: for i #= j and a e [0,1], The ramps are created by changing a, so that if one chooses a ramp s, s,- then the result is observations from the set:

[0272] { s(i,j, a) : a e linspace(0, 1, / V)} where N « 30 for all ramps. For each data-point a two dimensional output is obtained which corresponds to solvent concentration and product concentration, both of these are combined to calculate the yield which will be the target of the predictive models investigated. It is possible to parameterise each of the solvents, s, using a 13 dimensional representation %(s) e ]R13which is known for each of the base solvents. The assumption is made that a representation of any solvent can be obtained in the search space by considering the linear combination of the representation, i.e.: x(s i,j, a)) = a x st) + (1 — a) x(Sj)

[0273] The goal will be to understand the space of yield over solvents, i.e.:

[0274] Note that this example is learning in a one-dimensional input subspace since each ramp is one-dimensional in a and the aim is to investigate the union of such spaces. This makes the problem tractable.

[0275] Search space reduction based on data

[0276] The collected data corresponds to 9 ramps. The full space corresponds to 55 different combinations of solvents, therefore less than 20% of the space of solvents is covered. In order to investigate the suitability of Bayesian data collection algorithms, this example focuses on maximizing the predictive performance of out models only on this small subspace. This is to provide proof of concept that the algorithm can provide value when trying to learn the whole space. Fitting a Gaussian Process

[0277] The fitting of standard Machine Learning models was investigated: linear regression, random forest, decision tree and support vector machines. The SVM had the best performance. It was decided to investigate the fitting of a Gaussian Process model - this is because the aim is to leverage the uncertainty estimates they provide to be able to use efficient data collection algorithms. Chosen was a Gaussian process (GP) with a sparse axis-aligned subspace prior (Eriksson, David, and Martin Jankowiak. "High-dimensional Bayesian optimization with sparse axis-aligned subspaces." Uncertainty in Artificial Intelligence. PMLR, 2021) since the GP is to be fit to 13-dimensional features. What this GP does is favour longer length-scales effectively ‘turning off’ certain dimensions if it does not deem them important; This has shown to be effective in higher dimensional modeling. The standard choice of a Matern 5-2 kernel was used.

[0278] Cross-validation analysis

[0279] In this section was tested the suitability of Gaussian processes to predict the solvent ramps. To do this, carried out was a leave one out cross validation procedure and obtained was the following predictions for each of the 9 ramps, resulting in the predictions shown in Figure 13.

[0280] Overall the model seems to learn enough patterns to have generally accurate predictions in most folds. However, the real encouraging aspect of the results is that a variety of confidence intervals can be seen which appear to accurately represent the model’s uncertainty, as most data-points have the real values falling inside the credible intervals. This implies that the uncertainty quantification can be used by data collection algorithms for selecting the best ramps.

[0281] Using Gaussian processes for data collection

[0282] In this section an illustrative example is provided of how to use Gaussian processes and the corresponding uncertainty estimates to make decisions. Consider the situation where a GP model was trained on all ramps except two, and one wishes to make a decision of which of the two ramps to add to the data-set. For example, a GP was trained on the data from all ramps except for water methanol and HFIP -> I PA.

[0283] The experimental predictions for the two missing ramps was obtained as shown in Figure 14. It can be see that the model is very certain of the predictions on one of the ramps and very uncertain of its predictions on the second graph. From this we can conclude that if we were to gather one more set of experiments, HFIP IPA should be chosen as to minimize the model uncertainty.

[0284] Formalizing the arguments via information theory

[0285] In the previous section was an illustrative example that gives an intuition for how Bayesian data gathering algorithms work. In this section an algorithm is explained that makes the intuitive ideas into a concrete procedure.

[0286] The concept introduced in the previous section is formalised using information theory. This was done by following the ideas from the sensor placing algorithm introduced in Krause et al. (2008) (Krause, Andreas, Ajit Singh, and Carlos Guestrin. "Near-optimal sensor placements in Gaussian processes: Theory, efficient algorithms and empirical studies." Journal of Machine Learning Research 9.2 (2008)). In that work, when selecting a sensor location they propose choosing the place which will maximize the mutual information (Ml) of target set of locations U. In particular, given a set of locations already chosen A, a set of target locations U, a set of possible locations S, and V = U u S. For our case, we set V = U = S to be the set of points in the 9 ramps in the Claisen data-set. Then the new sensor location x* is chosen by maximizing the increase in mutual information gain:

[0287] MI(A U x) - M / Q4) = H x |A) - H(x\A) which is equivalent to maximizing the objective: where X is the covariance matrix and A = V \ {A U } (i.e. the set of possible locations and locations of interest but without the set of sensors included already). This can be extended to consider batches of points, X = (x1;. . . , x„). In particular, letting A = y \ ( A u X) allows selection of the best batch of points if one optimises the objective t7(X):

[0288] [ / (X) = MI(A U X) - MI(A) = H(X\A) - H(X\A)

[0289] The conditional entropies can be classically be calculated by the log determinant of the posterior covariance matrix, therefore giving a criterion:

[0290] In the data-set there was sometimes singular covariance matrices due to repeated measurements. Therefore the pseudo-determinants were used instead, which for semi-positive definite matrices (which is applies in this use case) are calculated as the product of the non-zero singular values of the matrix, due to possible numerical imprecisions, we deemed a singular value to be zero if it was smaller than 10-3. Since each ramp r e 5? consists of a batch of input locations Xr, one can choose which ramps to query using the decision rule: rn+1= argmaxre3J[ / (r) where the optimization sub-problem can be solved via enumeration due to the low number of ramps considered. It is noted that there is an implicit dependence on the hyper-parameters of the GP, 0, which defines the covariance function. This dependence is made explicit by writing [ / (r|6).

[0291] Data gathering experiment

[0292] Finally, in this section the algorithm is benchmarked on the data-set. The goal is to gather as much information across the whole space of 9 ramps, therefore we measure algorithmic performance via Mean Squared Error (MSE) across all ramps. Ramps were added iteratively and it was seen how MSE improves, comparing adding ramps via the mutual-information algorithm against adding ramps randomly.

[0293] Obtained was the GPs hyper-parameter by training on the whole set in a Bayesian manner, using a No U-Turn Sampling (NUTS) Monte Carlo algorithm, and obtaining a sample of size 32; this allows one to define an approximate objective after marginalization over the hyperparameters: where each of the 0 ~ p(0 |Z>) is a Monte Carlo sample.

[0294] The algorithm was run as described in the section above (“Formalizing the arguments via information theory”) for 9 iterations, giving the following preferred order for how the ramps should be gathered: acetic acid acetonitrile ethanol water HFIP -> IPA 2-methyl THF ethylene glycol ethylene glycol water cyclohexane ethanol methanol ethylene glycol ethyl acetate methanol methanol water The method was tested by measuring the MSE of the model after training on each of the ramps. At each iteration a new ramp was added and the MSE calculated on the whole data-set. As a baseline, the procedure was repeated with 50 random shuffles and the same test done, and the median and inter-quantile ranges plotted. The results are shown in Figure 15.

[0295] Overall by using the Ml-based selection criteria the performance is never worse than adding ramps randomly, and there was very strong performance from the selection of two ramps onwards. The performance of the algorithm is very encouraging given the small space of the search space, where one would expect a random selection to be powerful.

[0296] Conclusion

[0297] In this example Bayesian modelling of solvent ramp data was looked at, and the suitability of Gaussian processes for data acquisition was investigated within this task. Even with a limited data-set, these results show the promise of information-seeking data collection policies, which can have a much larger impact when choosing which experiments to run in a larger search space.

Claims

Claims1. A method for monitoring of one or more process parameters of a chemical process to convert one or more input substances to one or more output substances at different solvent ratios, comprising: a) providing a plurality of input streams, each comprising a solvent or solvent mixture and one or more input substances for the chemical process, wherein at least two of the input streams comprise a distinct solvent or solvent mixture; b) delivering one or more of the input streams to a flow reactor; c) varying the ratio of solvents within the flow reactor by changing relative flow rates of the input streams to the flow reactor; and d) monitoring one or more process parameters at different solvent ratios.

2. A method according to claim 1, wherein varying the ratio of solvents comprises a continuous linear or non-linear ramping of the relative flow rates.

3. A method according to claim 1 or claim 2, wherein the chemical process is selected from a chemical reaction, an extraction, a crystallisation, and a quenching step.

4. A method according to any one of claims 1 to 3, wherein the method is for concurrently monitoring one or more process parameters of multiple chemical processes, wherein for step (a) the input substances for each chemical process are provided in the plurality of input streams, and step (d) comprises monitoring one or more process parameters for each chemical process at different solvent ratios, wherein each input stream comprises the input substances for all of the chemical processes, or each input stream comprises the input substances for one chemical process.

5. A method according to any one of the preceding claims, wherein the method is quantitative and the one or more process parameters are quantitative parameters.

6. A method according to any one of the preceding claims, wherein in step (c) the ratio of solvents is varied whilst maintaining a constant total flow rate to the flow reactor.

7. A method according to any one of the preceding claims, wherein the chemical process is a chemical reaction to produce one or more products from one or more starting materials, wherein the one or more input substances are one or more starting materials and the one or more output substances are the one or more products, and the method comprises (e)determining reaction performance from the monitored one or more process parameters at different solvent ratios.

8. A method according to claim 7, wherein the one or more process parameters are selected from: product yield, reaction species concentration, heat of reaction, weight average molecular weight, number average molecular weight, gas offtake, or gas uptake, purity, enantiomeric excess, dispersity, selectivity, rate of reaction, reaction orders, activation energy, or combinations thereof.

9. A method according to any one of the preceding claims, comprising three or more input streams, wherein three of the input streams comprise a distinct solvent or solvent mixture.

10. A method according to any one of the preceding claims, wherein the chemical process is a unimolecular, bimolecular or multicomponent chemical reaction.

11. A method according to any one of claims 1 to 10, wherein the chemical process is a chemical reaction selected from a homogeneous organic chemical reaction, a heterogenous catalysis reaction, a homogeneous catalysis reaction, a biphasic liquid-gas reaction, a nano particle synthesis reaction, a crystallisation reaction, a biological reaction, a fermentation reaction, or a polymerisation reaction.

12. A method according to any one of the preceding claims, wherein the solvent or solvent mixtures comprise a liquid at room temperature and pressure.

13. A method according to any one of the preceding claims, wherein the flow reactor is a gas-liquid contactor.

14. A method according to any one of the preceding claims, wherein the solvent or solvent mixtures of the plurality of input streams comprise a supercritical fluid.

15. A method according to any one of the preceding claims, wherein the solvent or solvent mixtures of the plurality of input streams comprise a deep eutectic solvent and / or inorganic derived solvent.

16. A method according to any one of the preceding claims, wherein the solvent or solvent mixtures of the plurality of input streams comprise an organic solvent selected from one or more of polar, apolar, protic, aprotic, halogenated, aqueous, and aromatic solvents.

17. A method according to any one of the preceding claims, wherein the one or more process parameters are monitored by in-line or on-line monitoring.

18. A method according to any one of claims 1 to 16, wherein the one or more process parameters are monitored by at-line or off-line monitoring.

19. A method according to any one of the preceding claims, wherein the one or more process parameters are monitored by using High-performance Liquid Chromatography (HPLC).

20. A method according to claim 19, wherein the one or more process parameters are monitored using normal phase, reverse phase, chiral, or size exclusion HPLC.

21. A method according to claim 19 or claim 20, wherein the HPLC comprises one or more of a diode array detector (DAD), refractive index detector (RID), mass spectrometry detector (MSD), or evaporative light scattering detector (ELSD).

22. A method according to any one of claims 1 to 18, wherein the one or more process parameters are monitored by using one or more spectroscopic or spectrometric tools, either on its own or in combination with chemometric analysis.

23. A method according to claim 22, wherein the tool is selected from the list of UV / Vis, FTIR, NearIR, Raman, X-ray, fluorescence, NMR, or Mass Spectrometry.

24. A method according to any one of claims 1 to 18, wherein the one or more process parameters are monitored by using thermography, flow calorimetry, or gas chromatography.

25. A method of training a machine learning model to predict one or more process parameters of a chemical process to convert one or more input substances to one or more output substances at different solvent ratios of at least two solvents, or for a solvent, the method comprising: monitoring the one or more process parameters at different solvent ratios of one or more pairs of solvents using the method of any preceding claim; recording training data pairs of respective solvent ratio data points and associated process parameter data points, the solvent ratio data points comprising a value of solvent ratioand the process parameter data points comprising respective one or more values of the one or more process parameters; and training a machine learning model using the training data pairs to output a predicted process parameter data point in response to an input of a solvent ratio data point.

26. The method of claim 25, wherein the machine learning model outputs a measure of uncertainty for each predicted process data point.

27. The method of claim 26, wherein the machine learning model is a Gaussian Process.

28. The method of claim 26 or 27, wherein the measure of uncertainty is or is derived from a variance or a covariance matrix.

29. The method of claim 26, 27 or 28, comprising: generating a respective first data set for each of at least two further pairs of solvents, each first data set comprising, for the respective pair of solvents, predicted process parameter data points output by the trained machine learning model for respective solvent data points and a respective measure of uncertainty output by the trained machine learning model associated with each predicted process parameter; selecting at least one of the at least two further pairs of solvents based on the measures of uncertainty in the first data sets; for each of the selected at least one further pair of solvents, monitoring the one or more process parameters at different solvent ratios of a pair of solvents using the method of any one of claims 1 to 24 and recording further training data pairs of respective solvent ratio data points and associated process parameter data points; and training the machine learning model using the further training data pairs to output a predicted process parameter data point in response to an input of a solvent ratio data point.

30. The method of claim 29, wherein selecting at least one of the at least two further pairs of solvents based on the measure of uncertainty comprises selecting at least one of the at least two further pairs of solvents based on a measure of mutual information, optionally local mutual information, for each first data set.

31. The method of claim 30, comprising predicting a respective second data set for each of the one or more pairs of solvents, each second data set comprising, for the respective pair of solvents, predicted process parameter data points for respective solvent data points and a respective measure of uncertainty associated with each predicted process parameter,wherein the measure of mutual information is a respective gain in mutual information when each first data set is added to the second data set.

32. The method of claims claim 31 , wherein the pair of solvents that has the largest measure of mutual information associated with it is selected.

33. The method of any one of claims 29 to 32: determining if a stopping criterion is met; and if the stopping criterion is not met, repeating the generating, selecting, monitoring, recording and training.

34. The method of claim 33, wherein the stopping criterion comprises a measure of an error in the prediction by the machine learning model being less than or equal to a threshold value.

35. The method of claim 33, when dependent on claim 30, 31 or 32, wherein the stopping criterion comprises the measure of mutual information being less than or equal to a threshold value.

36. A method of designing a chemical process to convert one or more input substances to one or more output substances, the method comprising using a machine learning model trained according to the method of any one of claims25 to 35 to predict respective one or more process parameters for a plurality of ratios of solvents of a pair of solvents, or for a solvent; selecting a solvent ratio based on the predicted respective one or more process parameters; selecting the solvents of the pair of solvents at the selected solvent ratio, or a solvent, for the process.

37. The method of claim 36 comprising using the machine learning model to predict respective one or more process parameters for a plurality of ratios of solvents of a pair of solvents for each of a plurality of pairs of solvents; selecting a solvent ratio and pair of solvents based on the predicted respective one or more process parameters; selecting the solvents of the selected pair of solvents at the selected solvent ratio, or selected solvent, for the process.

38. A system for monitoring one or more process parameters of a chemical process to convert one or more input substances to one or more output substances at different solvent ratios, comprising:(i) a plurality of variable flow pumps, each in fluid communication with an input stream comprising a solvent or solvent mixture and the one or more input substances for the chemical process, wherein at least two of the input streams comprise a distinct solvent or solvent mixture;(ii) a flow reactor in fluid communication with each of the variable flow pumps; and(iii) an analytical unit in fluid communication with the flow reactor, configured to monitor one or more process parameters.

Images