# Fileset

[d3dd00215b.pdf](https://mdr.nims.go.jp/filesets/67de3765-19bd-4b66-a086-5b38ae2c96c2/download)

## Creator

Yota Fukui, [Kosuke Minami](https://orcid.org/0000-0003-4145-1118), [Kota Shiba](https://orcid.org/0000-0001-7775-0318), [Genki Yoshikawa](https://orcid.org/0000-0002-9136-8964), [Koji Tsuda](https://orcid.org/0000-0002-4288-1606), [Ryo Tamura](https://orcid.org/0000-0002-0349-358X)

## Rights

[Creative Commons BY Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/)

## Other metadata

[Automated odor-blending with one-pot Bayesian optimization](https://mdr.nims.go.jp/datasets/13253e9a-cfd2-4f40-999a-03f6ef2bd606)

## Fulltext

Automated odor-blending with one-pot Bayesian optimizationDigitalDiscoveryPAPEROpen Access Article. Published on 16 April 2024. Downloaded on 6/7/2024 9:19:13 PM.  This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.View Article OnlineView Journal  | View IssueAutomated odoraGraduate School of Frontier Sciences, TheKashiwa, Chiba 277-8568, Japan. E-mail: taac.jpbCenter for Basic Research on Materials, NaNamiki, Tsukuba, Ibaraki 305-0044, JapancResearch Center for Macromolecules anMaterials Science, 1-1 Namiki, Tsukuba, Ibkosuke@nims.go.jpdMaterials Science and Engineering, GraduUniversity of Tsukuba, 1-1-1 Tennodai, Tsuk† Electronic supplementary informahttps://doi.org/10.1039/d3dd00215bCite this: Digital Discovery, 2024, 3,969Received 31st October 2023Accepted 3rd April 2024DOI: 10.1039/d3dd00215brsc.li/digitaldiscovery© 2024 The Author(s). Published by-blending with one-pot Bayesianoptimization†Yota Fukui,ab Kosuke Minami, *c Kota Shiba, c Genki Yoshikawa, cdKoji Tsuda *ab and Ryo Tamura *abThe creation of new odors by blending existing ones is usually donemanually based on the human sense. Toenable robots to perform this automatically, we developed an automated odor-blending system. In thissystem, an olfactory sensor system composed of an array of Membrane-type Surface stress Sensors(MSSs) performs odor measurement of a blended liquid, and Bayesian optimization controls the blendingconcentration. The actual blending of the liquid samples is performed by automated syringe pumps. Oursystem performs odor-blending by injecting liquid samples into a pot or by draining some of the liquidfrom the pot. The one-pot strategy has the advantage of reducing the amount of liquid samples used inthe entire optimization task and reduces the problem of pot replacement. To implement this one-potstrategy effectively, a Drainable One-Pot Bayesian Optimization (DOPBO) algorithm was developed andapplied to our system. The system was tested using a ternary liquid mixture.IntroductionOdor is a complex gas mixture consisting of thousands ofdifferent molecules. Humans perceive different odors when thecomposition and concentration of the gas molecules arealtered. Since there are over 400 000 types of odorous/odorlessmolecules,1,2 an exceptionally large number of odors exist.Although such complexity of odors makes it difficult to createthe desired odors, avor chemists can create odors by blendingnatural and articial fragrances based on the human sense.3,4However, the human sense for odor varies signicantly acrossindividuals. Moreover, the number of avor chemists is limitedbecause of the rigorous training required to become one.If we replace the avor chemists with automated odor-blending systems, even with a limited number of avor chem-ists, a desired odor can be created conveniently by automaticallyblending a small number of odor samples. Aer the blendrecipe of a desired odor is identied, the odor can be digitized,and its odor information can be shared worldwide. In addition,University of Tokyo, 5-1-5 Kashiwanoha,mura.ryo@nims.go.jp; tsuda@k.u-tokyo.tional Institute for Materials Science, 1-1d Biomaterials, National Institute foraraki 305-0044, Japan. E-mail: minami.ate School of Pure and Applied Science,uba, Ibaraki 305-8571, Japantion (ESI) available. See DOI:the Royal Society of Chemistryit would be feasible to handle the cases where the blending ofodors is difficult by human from a safety perspective. That is, itwould be possible to create odors from hazardous samples,chemicals, and gases for the human body by blending safe odorsamples. This would enable humans to understand thesehazardous odors and dramatically reduce the number of acci-dents based on hazardous odors such as natural gas leakedfrom pipelines5 and volcanic gases.6 To establish an automatedodor-blending system, we need to develop three elements: anolfactory sensor7–13 to replace the human sense of odors, black-box optimization14–16 to determine the amount of blending, anda robotic system17–25 to perform actual blending of samples.In this study, we integrated the olfactory sensor technique,black-box optimization method, and automated devices todevelop a robotic system that can automatically blend odors ina one-pot system (Fig. 1). Here, the blending of the liquidsamples was targeted. To produce a liquid mixture that exhibitsthe desired odor, liquid samples are injected into a pot. Some ofthe mixed liquid sample in the pot can be drained. The advan-tages of our one-pot system are as follows: (i) the amount of themixed liquid sample in a pot does not vary signicantly, (ii) theamount of liquid samples used in the entire optimization taskcan be reduced, and (iii) the problem of replacing the pot iseliminated. The odor of the mixed liquid sample is a gas thatevaporates from the sample and is measured using an olfactorysensor. Variations in the amount of the mixed liquid sample inthe pot can affect the concentration of gas in the headspace,26,27and the response of the olfactory sensor should change. Thus,maintaining a constant amount of the sample in the pot is animportant factor for stable measurements by the olfactory sensorsystem. However, if we prepare mixed liquid samples in differentDigital Discovery, 2024, 3, 969–976 | 969http://crossmark.crossref.org/dialog/?doi=10.1039/d3dd00215b&domain=pdf&date_stamp=2024-05-11http://orcid.org/0000-0003-4145-1118http://orcid.org/0000-0001-7775-0318http://orcid.org/0000-0002-9136-8964http://orcid.org/0000-0002-4288-1606http://orcid.org/0000-0002-0349-358Xhttps://doi.org/10.1039/d3dd00215bhttp://creativecommons.org/licenses/by/3.0/http://creativecommons.org/licenses/by/3.0/https://doi.org/10.1039/d3dd00215bhttps://pubs.rsc.org/en/journals/journal/DDhttps://pubs.rsc.org/en/journals/journal/DD?issueid=DD003005Fig. 1 (a) Overview of the automated odor-blending system using theMSS and DOPBO. DOPBO indicates the next injection amounts ofliquid samples and the discharge of drainage to resemble the targetresponse signals. The similarity between the measured and targetsignals for the four selected channels is calculated. This is used to trainthe Gaussian process regression in DOPBO. The channel selection isperformed to correctly predict the concentration of the mixture. (b)Photograph of the automated odor-blending system. Our system isconstructed with MSS module, syringe pumps, a pot with the mixedsample, and mass flow controllers.Digital Discovery PaperOpen Access Article. Published on 16 April 2024. Downloaded on 6/7/2024 9:19:13 PM.  This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.View Article Onlinepots in each optimization cycle, the amount of liquid samplesrequired to complete the optimization would increase, and thereplacement of the pot in the odor measurement system wouldbe time-consuming. The one-pot strategy can solve these prob-lems simultaneously. We developed a Bayesian optimization(BO) algorithm for a one-pot odor-blending system, which iscalled the Drainable One-Pot Bayesian Optimization (DOPBO)algorithm. An olfactory sensor system composed of an array ofMembrane-type Surface stress Sensors (MSSs)28–30 with 12 chan-nels was used. The robot system was developed by automatingthe operation of a syringe pump using an in-house LabVIEWprogram. In our system, the response signals of the MSS for thetarget odor were rst measured. To reproduce the responsesignals, the liquid samples were mixed in a one-pot system. Oursystem was tested to create the odors of a ternary liquid mixtureand two seasonings.970 | Digital Discovery, 2024, 3, 969–976MethodsOne-pot Bayesian optimization algorithm (OPBO)Before introducing the BO algorithm used in our one-pot odor-blending system, called DOPBO, we explain the algorithmwithout drainage, which is called One-Pot Bayesian Optimiza-tion (OPBO). This method optimizes the concentration ofa liquid mixture. It minimizes the objective function f(w) bysuccessively injecting the liquid samples into a pot. In oursystem, the objective function f(w) is dened by the similarity ofresponse signals between the target odor and the liquid mixturein the pot. When three liquid samples are to be mixed, theOPBO procedure is as follows:(i) If there are M mixed liquid samples for which the objec-tive function f(w) with the concentration of w is known inadvance, the initial training data are dened as Dtrain = {(wi,f(wi))}i=1,.,M. When the amount of the three liquid samples isdened as x = (x1, x2, x3), the concentration of each liquidsample w = (w1, w2, w3) is obtained usingwi ¼ xix1 þ x2 þ x3; ði ¼ 1; 2; 3Þ: (1)(ii) The initial amounts of the three liquid samples in the potare determined as x1 = (x11, x12, x13). The correspondingconcentrations are calculated using eqn (1), and the objectivefunction of f(w1) is measured. These data are added to Dtrain,and the number of training data becomes M + 1.(iii) We assume that (d1, d2, d3) are the minimum units of theinjection amounts for each liquid sample and the total injectionamounts should not exceed D. Using these parameters, weprepare a dataset Dcand that lists the amounts of the three liquidsamples in the pot aer injecting liquid samples. For example,when d1 = d2 = d3 = 0.1 and D = 0.2, the candidate dataset fromx1 = (x11, x12, x13) is dened asDcand = {(x11, x12, x13 + 0.1), (x11, x12 + 0.1, x13),(x11 + 0.1, x12, x13), (x11, x12, x13 + 0.2),(x11, x12 + 0.2, x13), (x11 + 0.2, x12, x13),(x11, x12 + 0.1, x13 + 0.1), (x11 + 0.1, x12 + 0.1, x13),(x11 + 0.1, x12, x13 + 0.1)}. (2)(iv) We prepare a candidate dataset containing the concen-trations Dcandw. These were converted from Dcand using eqn (1).(v) The Gaussian process regression is trained using Dtrain.The objective function and its uncertainty are predicted whenDcandw is used as the testing dataset. A promising candidatethat minimizes the objective function is selected using anacquisition function based on the predicted value and itsuncertainty.(vi) The liquid samples are injected according to the selectedcandidate. As a result, the amounts of the three liquid samplesin the pot become x2 = (x21, x22, x23). The objective function ismeasured as f(w2) with the concentration of w2 converted fromx2. These data are added to Dtrain, and the number of trainingdata becomes M + 2.(vii) OPBO can be performed by repeating steps (iii)–(vi).© 2024 The Author(s). Published by the Royal Society of Chemistryhttp://creativecommons.org/licenses/by/3.0/http://creativecommons.org/licenses/by/3.0/https://doi.org/10.1039/d3dd00215bFig. 2 Schematic image of the experimental environment.Paper Digital DiscoveryOpen Access Article. Published on 16 April 2024. Downloaded on 6/7/2024 9:19:13 PM.  This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.View Article OnlineIn this study, the Bayesian optimization package PHYSBO31was used in (v), and Thompson sampling is used to generate theacquisition function. The OPBO algorithm has no upper limit ofthe number of liquid samples although the case of mixing threeliquid samples is used as an example.In the BO approach, it is common to use a xed objectivefunction of f(w). On the other hand, in our study, we adopted thestrategy where the objective function is updated in each cycle.This is because it is not known in advance which of the multipleMSS channels is effective for the target odors. In this case, in(vi), the effective channels are appropriately selected usinga method to be explained later, and f(w) is updated to the newobjective function dened by the selected channels andreturned to step (iii).Drainable one-pot Bayesian optimization algorithm (DOPBO)In the OPBO algorithm, only the injection into the pot isconsidered. However, if we can consider the drainage from thepot, the search space of the liquid mixture can be extendedfurther. In addition, since the amount of liquid sample in thepot is not only increasing, the amount of sample in the potwould remain constant. DOPBO is a drainable version of OPBO.In this algorithm, in the step (iii) of OPBO, the number ofcandidates increases because of the diverse concentrationsowing to drainage. Let h= (h1, h2,.,hn) be the candidates of theratio of the amounts to be reduced from the pot. For example, hi= 0 represents the case where no liquid sample is drained, andhi = 0.4 represents the case where 40% of the liquid sample inthe pot is drained. When d1 = d2 = d3 = 0.1 and D = 0.2, thecandidate dataset is dened asDcand = {[x11 × (1 − hi), x12 × (1 − hi), x13 × (1 − hi) + 0.1],[x11 × (1 − hi), x12 × (1 − hi) + 0.1, x13 × (1 − hi)],[x11 × (1 − hi) + 0.1, x12 × (1 − hi), x13 × (1 − hi)],[x11 × (1 − hi), x12 × (1 − hi), x13 × (1 − hi) + 0.2],[x11 × (1 − hi), x12 × (1 − hi) + 0.2, x13 × (1 − hi)],[x11 × (1 − hi) + 0.2, x12 × (1 − hi), x13 × (1 − hi)],[x11 × (1− hi), x12 × (1 − hi) + 0.1, x13 × (1− hi) + 0.1],[x11 × (1− hi) + 0.1, x12 × (1 − hi) + 0.1, x13 × (1− hi)],[x11 × (1 − hi) + 0.1, x12 × (1 − hi), x13× (1 − hi) + 0.1]}i=1,.,n. (3)Compared with OPBO, the search space increases, and moreaccurate optimization may be performed. DOPBO is availableon GitHub (https://github.com/tsudalab/DOPBO). In the script,physbo 1.1.1 is used.Automated odor-blending system for a ternary liquid mixtureAs shown in Fig. 2, four syringe pumps were set up. The error inadding liquid samples with syringe pumps was approximately 3microliters. Three of them contained three types of liquidsamples, and the fourth was empty. The sample was supplied at5 mL min−1 by connecting PTFE tubes to the syringes. Thesample was added dropwise by inserting the PTFE tubesthrough the holes in the lid of glass vial 1. A PTFE tube con-nected to the empty syringe was also inserted into a hole in the© 2024 The Author(s). Published by the Royal Society of Chemistrylid of vial 1 so that the mixed liquid sample in vial 1 could bedrained. Vial 1 was maintained at 25 °C using a heating bath.The mixed liquid sample was stirred well with a small magneticstirring bar at 300 rpm. Vial 1 was connected to two PTFE tubes.One of them was connected to mass ow controller 1, and 40sccm of nitrogen was supplied to vial 1. The other PTFE tubewas connected to the sampling gas inlet of the standard modulecontaining the MSS chips with 12 channels and was set tocapture the gas in vial 1 at 30 sccm. The nitrogen supply line wascreated using mass ow controller 1 to prevent measurementerrors caused when outside air with high humidity is drawn in.Specically, by providing a larger amount of nitrogen to vial 1than the sampling inow from vial 1, we prevented the outsideair from entering its headspace (the space with only the targetgas above the mixed liquid sample). Mass ow controller 2 wasconnected to the purge gas inlet of the standard modulethrough empty glass vial 2 and supplied nitrogen at 40 sccm.Moreover, the purge gas inlet of the standard module was set tocapture gas at 30 sccm. This difference between the two owrates was maintained to prevent the outside air with highhumidity from entering vial 2.To realize an automated odor-blending system, the LabVIEWprogram was developed to control the pumping by the syringepumps, perform DOPBO, and analyze the response signals of theMSS. Our developed LabVIEW program (National Instruments)and the program controlling the standard module for the MSSwere executed simultaneously on the same PC. Using the Lab-VIEW program, the timing for measuring the target sample bythe MSS was specied, and a demonstration was conducted withmeasurements at approximately 8 min intervals. The devicesused in our system are summarized in ESI Note A.†Features of each response in a signalTo obtain the response signals for the odor sample, thefollowing protocol using MSSs was iterated. First, the samplingand purging were repeated two times for 10 s each. Subse-quently, a long purge time of 120 s was applied. A schematic ofthis response is shown in Fig. 3. For the second peak with a 10 ssample and purge, four parameters were extracted. Parameter 1is dened by the slope of the line joining points A and B shownin Fig. 3. This is related to the adsorption process on the odorreceptor materials. The slope of the line joining points B and Cwas used as parameter 2, and the quasi-equilibrium stateDigital Discovery, 2024, 3, 969–976 | 971https://github.com/tsudalab/DOPBOhttp://creativecommons.org/licenses/by/3.0/http://creativecommons.org/licenses/by/3.0/https://doi.org/10.1039/d3dd00215bFig. 3 Schematic response signal from the MSS and the protocol toobtain the feature of the response signal.Digital Discovery PaperOpen Access Article. Published on 16 April 2024. Downloaded on 6/7/2024 9:19:13 PM.  This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.View Article Onlineinformation was reected. Parameter 3 is dened as the slope ofthe line joining points C and D. It is related to the odordesorption process. Finally, parameter 4 represents the heightof the signal and includes information on the adsorptioncapacity of each receptor material. Here, we xed the timedifference between points A and B as 0.5 s. It was also used forpoints C and D. Since we prepared 12 channels, each odorsample was represented by 48 parameters (i.e., four featurestimes 12 channels). That is, the feature vector extracted from theMSS signal was 48-dimensional. This denition of the featurewas used in our previous studies,32–34 and the odor quantica-tion and detection of quasi-primary odors can be achieved usingthese features.Selection strategy of effective channels in MSSsFrom the response signals obtained by the standard module,48-dimensional features were obtained. The objective functionfor DOPBO is dened by the similarity between the features ofthe target odor and liquid mixture. However, since effectivechannels depend on liquid samples, ineffective channels maybecome noise for the optimization task. The effective channelwill also be affected by the combination of samples to be mixed.Thus, we consider the case that the effective channel is notknown before optimization, and when the data are increased bythe optimization task, the search for the effective channel isperformed simultaneously with the optimization. We intro-duced a strategy in which four effective channels were selectedduring the optimization process, and the objective function wasdened using only the selected channels. Here, we assumedeffective channels to be those that can correctly predict theconcentration of the mixture. For each channel, four-dimensional features were extracted and used to predict theconcentration of each liquid sample by linear regression. Toestimate the prediction accuracy, we performed a ve-foldcross-validation and calculated the mean squared error foreach concentration of the mixture. The average value of themean squared errors for each concentration was used asa measure of the channel that could correctly predict theconcentration of the mixture. Thus, the top four channels inascending order of this value were selected as the effective972 | Digital Discovery, 2024, 3, 969–976channels. For each step in the optimization cycle, this selectionwas performed using the data obtained at that time. Since thedata for the initial step was inadequate, optimization wasstarted on a predetermined channel. When the number oftraining data points reached ve, the above selection wasstarted.Using the selected four effective channels, the objectivefunction for BO algorithms is dened asf(w) = jz(w) − z*j, (4)where z(w) and z* are the 16-dimensional feature vectors ob-tained from the four selected channels for the odor of the liquidmixture in the pot and target odor, respectively. When evaluatingeqn (4), the features were standardized to the training data.ResultsPerformance of drainable one-pot Bayesian optimizationusing a test functionWe evaluated the performance of DOPBO using a test functionbefore considering a real blending system. Assuming a ternarymixture case, a three-dimensional Ackley function was used.The target concentration for the three samples is dened asw* ¼ ðw*1;w*2;w*3Þ with 0#w*1 # 1; 0#w*2 # 1; 0#w*3 # 1 andw*1 þ w*2 þ w*3 ¼ 1. Rather than using eqn (4), the objectivefunction is set as follows when the concentration of the mixturein the pot is w:f ðwÞ ¼ 20� 20 exp8<:�0:2ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi13X3i¼1402�wi � w*i�2vuut9=;þ expð1Þ� exp"13X3i¼1cos�2p$40�wi � w*i��#: (5)This function is minimized when the concentrations of thetarget and mixture samples are equal, i.e., w = w*.To evaluate the optimization performance, 100 randomtarget concentrations w* ¼ ðw*1;w*2;w*3Þ were generated. Whenjw* − wj < 0.01 was achieved via optimization, the optimizationwas assumed to be successful. In the one-pot algorithms, theminimum unit of the injection amount for each liquid samplewas set to 0.1, i.e., d1 = d2 = d3 = 0.1. In addition, the maximumtotal injection amount at each step was set to D = 1.0. Thedependence on the minimum injection amount is discussed inESI Note B.† First, we considered the case where the initialtraining data included the data atwi= (1.0, 0.0, 0.0) and (0.0, 1.0,0.0), and the initial concentrations of the three liquid sampleswere set to w1 = (0.0, 0.0, 1.0). The number of targets thatattained jw* − wj < 0.01, where the optimization was succeeded,by the time of each step was counted. This value divided by 100 isdened as “success probability” and its step dependence isshown in Fig. 4(a). The results of conventional BO (where it isnecessary to empty the pot at every step), OPBOwithout drainage,and DOPBO were compared. Here, h = (0.0, 0.1, 0.2, 0.3, 0.4),which determines the amount to be reduced from the pot, wasused in DOPBO. That is, the maximum drainage was 40% of the© 2024 The Author(s). Published by the Royal Society of Chemistryhttp://creativecommons.org/licenses/by/3.0/http://creativecommons.org/licenses/by/3.0/https://doi.org/10.1039/d3dd00215bFig. 4 (a) Success probability for the 100 random targets, (b) totalamount used until each step, and (c) amount in the pot depending onthe step when conventional Bayesian optimization (BO), one-pot BO(OPBO), and drainable one-pot BO (DOPBO) with h = (0.0, 0.1, 0.2,0.3, 0.4) are used. The target function is the Ackley function. OPrandom and DOP random are the random exploration results whenthe same search spaces of OPBO and DOPBO are used, respectively.For the amounts, the average of the results for 100 random targets andthe standard deviation are shown as a line and shaded area.Paper Digital DiscoveryOpen Access Article. Published on 16 April 2024. Downloaded on 6/7/2024 9:19:13 PM.  This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.View Article Onlineliquid mixture in the pot. For BO, the minimum unit ofconcentration was xed as 0.01, and 5148 candidates weregenerated, and the amount of prepared mixture was assumed tobe one. One-pot versions of a random exploration were alsocompared as One-Pot (OP) random without drainage andDrainable One-Pot (DOP) random. In OP random and DOPrandom, from the candidate datasets dened by eqn (2) and (3),one condition is randomly selected, respectively.Fig. 4(a) shows that OPBO performed better than randomexplorations, with successful optimization of approximatelyhalf of the target concentrations aer 50 steps. The perfor-mances of BO and DOPBO were comparable. This veried thatoptimization was achieved at most of the target concentrations.These results highlight the importance of drainage in one-potsystems. The total amount of liquid samples used for each© 2024 The Author(s). Published by the Royal Society of Chemistrymethod depending on the steps is shown in Fig. 4(b). It wasobserved that the total amount of liquid samples could bereduced by DOPBO compared with the conventional BO. Theamount of liquid samples in the pot at each step is shown inFig. 4(c). The amount of liquid in the pot was controlled whendrainage was considered. This is an important aspect whenconducting experiments using olfactory sensors. It indicatesthat DOPBO is a suitable algorithm for automated odor-blending that can reduce the total amount of liquid samplesused compared with that for the conventional BO. In contrast,OPBO is not suitable in the absence of drainage because of itslow success probability and ineffective control over the amountof liquid in the pot. In ESI Note B,† the optimization perfor-mances depending on h, initial training dataset, and initialconcentrations are investigated. In addition, we performed thecalculations on several functions known as test functions with3-dimensional variables, and the results are summarized in ESINote C.† In all cases, we found that DOPBO performed betterand in some cases had a better success probability than BO.Demonstration of automated odor-blending in a real deviceTo demonstrate our automated odor-blending system(Fig. 1(b)), a mixture of 1-octanol (FujiFilmWako Pure ChemicalCorporation, 97%), isopropyl alcohol (FujiFilm Wako PureChemical Corporation, $98%), and methanol (Kanto ChemicalCo., Ltd, 99.8%) was considered. For the target liquid mixture,we prepared a sample of (1-octanol, isopropyl alcohol, meth-anol) = (0.2, 0.6, 0.2) and measured it using a MSS. Automatedodor-blending was performed using the initial training data of[(0.8, 0.1, 0.1), (0.1, 0.8, 0.1)], and (0.8 mL, 0.8 mL, 2.4 mL) as theinitial amount in the pot according to the best case for the testfunction shown in ESI Note B.† The initial concentrations of(0.2, 0.2, 0.6) are not close to the target concentration. Themaximum injection volume was set to D = 1.5 mL. Moreover, d1= d2 = d3 = 0.3 mL was used. The step dependences of theconcentrations of 1-octanol, isopropyl alcohol, and methanoland the value of the objective function f(w) dened by eqn (4)when the effective channels selected in the nal step were usedare summarized in Fig. 5(a). At the nal step, the value of theobjective function was minimized, and the concentrations ofthe mixture in the pot were close to the correct values in thiscycle. This result indicated that odor-blending was performedcorrectly using our robotic system. Fig. 5(b) shows the totalamount of liquid samples used in the optimization and thetransition of the amount of liquid in the pot. The amount ofliquid in the pot could be controlled. The selected effectivechannels depending on the step are shown in Fig. 5(c). Thisindicates that channels 2, 5, 10, and 11 are important in thenal step. These channels are different from the channels usedin the initial step, and the effective channels are selectedautomatically through the optimization cycle. The responsesignals obtained in the nal step for the selected channels at thenal step were compared with those obtained as the targets, assummarized in Fig. 5(d). The signals of the mixture accuratelyreproduced the target signals. In addition, an experiment whenthe target mixture is set to (1-octanol, isopropyl alcohol,Digital Discovery, 2024, 3, 969–976 | 973http://creativecommons.org/licenses/by/3.0/http://creativecommons.org/licenses/by/3.0/https://doi.org/10.1039/d3dd00215bFig. 5 (a) Concentrations of 1-octanol, isopropyl alcohol, and meth-anol depending on the step in the automated odor-blending system.The concentration of the target mixture was (1-octanol, isopropylalcohol, methanol) = (0.2, 0.6, 0.2). They are shown by the dottedlines. The value of the objective function f(w) is also shown when thechannels selected at the final step are used, i.e., channels 2, 5, 10, and11. (b) Total amount of liquid samples used in the optimization and theamount of liquid in the pot depending on the step. (c) Selected foureffective channels depending on the step. The blue points indicateused channels. (d) Comparison between signals obtained at the 8thstep and that obtained from the target mixture for the channelsselected at the final step. The similarity was evaluated in the pinkshaded area.Fig. 6 Concentrations of pure water, cooking sake, and fish saucedepending on the step in the automated odor-blending system when(a) mirin and (b) ponzu are targeted. The value of the objective functionf(w) is also shown when the channels selected at the final step areused. Comparison between signals of the mixture and those obtainedfrom the target for the channels selected at the final step when (c)mirin and (d) ponzu are targeted. The similarity was evaluated in thepink shaded area.Digital Discovery PaperOpen Access Article. Published on 16 April 2024. Downloaded on 6/7/2024 9:19:13 PM.  This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.View Article Onlinemethanol) = (0.6, 0.1, 0.3) was performed, and we conrmedthat the correct mixture was achieved. The results are summa-rized in Fig. S6.† Thus, we demonstrated that automated odor-blending can be realized correctly using the developed system.The operation of our automated odor-blending system is illus-trated in ESI Movie 1.†Finally, we attempted to create odors of some seasonings bymixing other seasonings using an automated odor-blendingsystem. We considered mixing pure water, cooking sake(Hinode Holdings Co., Ltd), and sh sauce (Allied CorporationCo., Ltd). These seasonings are known to have characteristicsignals measured by the MSS module,34 although it is differentfrom the current MSSmodule. The target seasonings were mirin(Hinode Holdings Co., Ltd) and ponzu (Mizkan Holdings Co.,Ltd). For each case, the step dependence of the concentrationsof pure water, cooking sake, and sh sauce and the value of theobjective function f(w) dened by eqn (4) when the effectivechannels selected in the nal step are used are summarized inFig. 6(a) and (b). In addition, the response signals at the step inwhich the objective function is minimized are compared withthose obtained as targets in Fig. 6(c) and (d). The total amountof liquid samples used in the optimization and the amount ofliquid in the pot and selected four effective channels dependingon the steps are summarized in Fig. S7.† In both cases, thesignals can be reproduced with high accuracy. For mirin, themain component in the mixture was cooking sake. This isconsistent with the fact that the main components of mirin are974 | Digital Discovery, 2024, 3, 969–976ethanol and some kinds of sugars, and the smell would besimilar to cooking sake. For ponzu, which is created by mixingsoy sauce, brewed vinegar, and citrus juice, the three seasoningsare mixed to form similar signals to target ones. On the otherhand, there is no sour smell in the mixed seasonings, and it isdifficult to perfectly reproduce the odor of ponzu based onvinegar and citrus, as a smell that human perceive.Discussion and conclusionsIn this study, we developed an automated odor-blending systemusing MSSs and machine learning. In our system, blending ofliquid samples was targeted, and the liquid samples were injec-ted into a pot. When olfactory sensors are used, it is important tocontrol the amount of liquid in the pot. That is, the amount ofliquid should not vary signicantly. To achieve this, an effectivealgorithm called drainable one-pot Bayesian optimization(DOPBO) was developed. Using some test functions, we demon-strated that the total amount of liquid used in the optimizationcan be decreased by DOPBO compared with conventionalBayesian optimization. Moreover, the optimization performanceof DOPBO is approximately equal to that of the conventionalBayesian optimization in the verication using test functions.Blending experiments were conducted using a mixture of 1-© 2024 The Author(s). Published by the Royal Society of Chemistryhttp://creativecommons.org/licenses/by/3.0/http://creativecommons.org/licenses/by/3.0/https://doi.org/10.1039/d3dd00215bPaper Digital DiscoveryOpen Access Article. Published on 16 April 2024. Downloaded on 6/7/2024 9:19:13 PM.  This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.View Article Onlineoctanol, isopropyl alcohol, and methanol. Liquid samples withthe correct concentration were produced successfully using theproposed system. In addition, we attempted to create odors fortwo seasonings (mirin and ponzu) by mixing other seasonings(pure water, cooking sake, and sh sauce) using our system andveried that mixing can produce sensor signals similar to thetarget signals. In the future, we will attempt to solve real prob-lems such as the blending of perfumes and articial avors.In this study, an MSS was used as the olfactory sensor. Thereceptor materials used in theMSS do not match human olfactoryreceptors, and the odors produced by our automated odor-blending systems are oen different from those perceived byhumans. The development of receptor materials with propertiessimilar to human olfactory receptors is an important perspectivefor the future, as it will complete the technology to create odorsthat are consistent with human perception. In addition, ourautomated odor-blending system can be easily applied to otherodor sensors such as metal oxide sensors,35,36 quartz resonatortype sensors,37 and piezoresistive sensors.38,39 Thus, the improve-ment of olfactory sensor technology will promote the practicalapplication of our automated odor-blending system.Data availabilityThe code of our algorithm called DOPBO and experimentalresults are available at https://github.com/tsudalab/DOPBO.Author contributionsYota Fukui: formal analysis (equal); methodology (equal); so-ware (equal); investigation (equal); visualization (equal); writing– original dra (equal). Kosuke Minami: methodology (equal);project administration (equal); writing – review & editing(equal). Kota Shiba: methodology (supporting); writing – review& editing (equal). Genki Yoshikawa: methodology (supporting);writing – review & editing (equal). Koji Tsuda: conceptualization(equal); methodology (equal); project administration (equal);writing – review & editing (equal). Ryo Tamura: conceptualiza-tion (equal); formal analysis (equal); methodology (equal);investigation (equal); soware (equal); visualization (equal);project administration (equal); writing – original dra (equal).Conflicts of interestThe authors declare no competing interests.AcknowledgementsWe thank Dissanayake Santha Kumara (NIMS) for the coating ofthe receptor layers onto the MSS chips. We thank TakahiroNemoto (NIMS), Masaaki Matoba (NIMS), and MasahitoKumada (U. Tokyo) for valuable discussions. This study waspartially supported by a project subsidized by a Grant-in-Aid forScientic Research (B), JSPS, MEXT, Japan (21H01971 and21H01008); a Grant-in-Aid for Scientic Research (C), JSPS,MEXT, Japan (22K05324); a Grant-in-Aid for ChallengingResearch (Exploratory), JSPS, MEXT, Japan (21K18859); Fund for© 2024 The Author(s). Published by the Royal Society of Chemistrythe Promotion of Joint International Research, InternationalCollaborative Research, JSPS, MEXT, Japan (JP19KK0141).References1 J. T. E. Richardson and G. M. Zucco, Psychol. Bull., 1989, 105,352–360.2 J. A. Gottfried, Curr. Opin. Neurobiol., 2009, 19, 422–429.3 H. Heath and G. Reineccius, Flavor Chemistry andTechnology, Springer Dordrecht, 1986.4 R. Teranishi, Challenges in Flavor Chemistry: An Overview,in Flavor Analysis, American Chemical Society, 1998, vol.705, pp. 1–6.5 A. Bianchini, A. Guzzini, M. Pellegrini and C. Saccani, Int. J.Pressure Vessels Piping, 2018, 168, 24–38.6 H. Sigurdsson, Encyclopedia of Volcanoes, Academic Press,2015.7 J. W. Gardner and P. N. Bartlett, Sens. Actuators, B, 1994, 18,210–211.8 K. Persaud and G. Dodd, Nature, 1982, 299, 352–355.9 G. Konvalina and H. Haick, Acc. Chem. Res., 2014, 47, 66–76.10 J. Gutiérrez and M. C. Horrillo, Talanta, 2014, 124, 95–105.11 R. A. Potyrailo, Chem. Rev., 2016, 116, 11877–11923.12 J.-W. Yoon and J.-H. Lee, Lab Chip, 2017, 17, 3537–3557.13 I. Manzini, D. Schild and C. Di Natale, Physiol. Rev., 2022,102, 61–154.14 T. Ueno, T. D. Rhone, Z. Hou, T. Mizoguchi and K. Tsuda,Mater. Discovery, 2016, 4, 18–21.15 K. Terayama, M. Sumita, R. Tamura and K. Tsuda, Acc. Chem.Res., 2021, 54, 1334–1346.16 Y. Jin and P. V. Kumar, Nanoscale, 2023, 15, 10975–10984.17 B. Burger, P. M. Maffettone, V. V. Gusev, C. M. Aitchison,Y. Bai, X. Wang, X. Li, B. M. Alston, B. Li, R. Clowes,N. Rankin, B. Harris, R. S. Sprick and A. I. Cooper, Nature,2020, 583, 237–241.18 A. Dave, J. Mitchell, K. Kandasamy, H. Wang, S. Burke,B. Paria, B. Póczos, J. Whitacre and V. Viswanathan, CellRep. Phys. Sci., 2020, 1, 100264.19 R. Shimizu, S. Kobayashi, Y. Watanabe, Y. Ando andT. Hitosugi, APL Mater., 2020, 8, 111110.20 B. P. MacLeod, F. G. L. Parlane, T. D. Morrissey, F. Häse,L. M. Roch, K. E. Dettelbach, R. Moreira, L. P. E. Yunker,M. B. Rooney, J. R. Deeth, V. Lai, G. J. Ng, H. Situ,R. H. Zhang, M. S. Elliott, T. H. Haley, D. J. Dvorak,A. Aspuru-Guzik, J. E. Hein and C. P. Berlinguette, Sci.Adv., 2020, 6, eaaz8867.21 L. M. Roch, F. Häse, C. Kreisbeck, T. Tamayo-Mendoza,L. P. E. Yunker, J. E. Hein and A. Aspuru-Guzik, PLoS One,2020, 15, e0229862.22 S. Matsuda, G. Lambard and K. Sodeyama, Cell Rep. Phys.Sci., 2022, 3, 100832.23 K. Nagai, T. Osa, G. Inoue, T. Tsujiguchi, T. Araki, Y. Kuroda,M. Tomizawa and K. Nagato, Sci. Rep., 2022, 12, 1615.24 R. Tamura, K. Tsuda and S. Matsuda, Sci. Technol. Adv.Mater.: Methods, 2023, 3, 2232297.25 K. Hippalgaonkar, Q. Li, X. Wang, J. W. Fisher, J. Kirkpatrickand T. Buonassisi, Nat. Rev. Mater., 2023, 8, 241–260.Digital Discovery, 2024, 3, 969–976 | 975https://github.com/tsudalab/DOPBOhttp://creativecommons.org/licenses/by/3.0/http://creativecommons.org/licenses/by/3.0/https://doi.org/10.1039/d3dd00215bDigital Discovery PaperOpen Access Article. Published on 16 April 2024. Downloaded on 6/7/2024 9:19:13 PM.  This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.View Article Online26 T. Górecki, Analyst, 1997, 122, 1079–1086.27 An Introduction to Headspace Sampling in GasChromatography, PerkinElmer, https://www.perkinelmer.com/libraries/gde_intro_to_headspace,accessed 13 February 2024.28 G. Yoshikawa, T. Akiyama, S. Gautsch, P. Vettiger andH. Rohrer, Nano Lett., 2011, 11, 1044–1048.29 G. Yoshikawa, T. Akiyama, F. Loizeau, K. Shiba, S. Gautsch,T. Nakayama, P. Vettiger, N. F. de Rooij and M. Aono,Sensors, 2012, 12, 15873–15887.30 K. Minami, G. Imamura, R. Tamura, K. Shiba andG. Yoshikawa, Biosensors, 2022, 12, 762.31 Y. Motoyama, R. Tamura, K. Yoshimi, K. Terayama, T. Uenoand K. Tsuda, Comput. Phys. Commun., 2022, 278, 108405.32 K. Shiba, R. Tamura, G. Imamura and G. Yoshikawa, Sci.Rep., 2017, 7, 1–12.976 | Digital Discovery, 2024, 3, 969–97633 K. Shiba, R. Tamura, T. Sugiyama, Y. Kameyama, K. Koda,E. Sakon, K. Minami, H. T. Ngo, G. Imamura, K. Tsudaand G. Yoshikawa, ACS Sens., 2018, 3, 1592–1600.34 H. Xu, K. Kitai, K. Minami, M. Nakatsu, G. Yoshikawa,K. Tsuda, K. Shiba and R. Tamura, Sci. Rep., 2021, 11, 12070.35 H. Meixner and U. Lampe, Sens. Actuators, B, 1996, 33, 198–202.36 C. Wang, L. Yin, L. Zhang, D. Xiang and R. Gao, Sensors,2010, 10, 2088–2106.37 K. Ema, M. Yokoyama, T. Nakamoto and T. Moriizumi, Sens.Actuators, 1989, 18, 291–296.38 Md. M. Hossain, M. Toda, T. Hokama, M. Yamazaki,K. Moorthi and T. Ono, IEEE Sens. Lett., 2019, 3, 1–4.39 M. Toda, K. Moorthi, T. Hokama, Z. Wang, M. Yamazaki andT. Ono, Sens. Actuators, B, 2021, 333, 129524.© 2024 The Author(s). Published by the Royal Society of Chemistryhttps://www.perkinelmer.com/libraries/gde_intro_to_headspacehttps://www.perkinelmer.com/libraries/gde_intro_to_headspacehttp://creativecommons.org/licenses/by/3.0/http://creativecommons.org/licenses/by/3.0/https://doi.org/10.1039/d3dd00215b Automated odor-blending with one-pot Bayesian optimizationElectronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3dd00215b Automated odor-blending with one-pot Bayesian optimizationElectronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3dd00215b Automated odor-blending with one-pot Bayesian optimizationElectronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3dd00215b Automated odor-blending with one-pot Bayesian optimizationElectronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3dd00215b Automated odor-blending with one-pot Bayesian optimizationElectronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3dd00215b Automated odor-blending with one-pot Bayesian optimizationElectronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3dd00215b Automated odor-blending with one-pot Bayesian optimizationElectronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3dd00215b Automated odor-blending with one-pot Bayesian optimizationElectronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3dd00215b Automated odor-blending with one-pot Bayesian optimizationElectronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3dd00215b Automated odor-blending with one-pot Bayesian optimizationElectronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3dd00215b Automated odor-blending with one-pot Bayesian optimizationElectronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3dd00215b Automated odor-blending with one-pot Bayesian optimizationElectronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3dd00215b Automated odor-blending with one-pot Bayesian optimizationElectronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3dd00215b Automated odor-blending with one-pot Bayesian optimizationElectronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3dd00215b Automated odor-blending with one-pot Bayesian optimizationElectronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3dd00215b Automated odor-blending with one-pot Bayesian optimizationElectronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3dd00215b