Screening for New Efficient and Sustainable-by-Design Solvents to Assist the Extractive Fermentation of Glucose to Bioethanol Fuels

Abstract

The production of bioethanol fuels using extractive fermentation increases the efficiency of the bioconversion reaction by reducing the toxic product inhibition. The choice of appropriate solvents to remove the bioethanol product without inhibiting the fermentation is important to enable industrial scale application. This work applies computer-aided molecular design technologies to systematically screen a wide variety of candidate solvents to enhance the separation, also considering the microorganisms that perform the fermentation. The performance of the candidates was evaluated using a rigorous process simulator for extractive fermentation, assisted by functional group-contribution (QSPR/QSAR) models for the prediction of various solvent properties, including toxicity and life cycle impacts. The solvent designs generated through this approach can provide powerful insights on the kind of molecular structures and functionalities that satisfy the process objectives and constraints, as well the desired sustainability features.

1. Introduction

The European Green Deal (EUGD) (COM(2019) 640 final) aims to make Europe the world’s first climate-neutral continent by 2050, and deliver benefits from sustainable green transition to European citizens and businesses [1]. EUGD raised the target for the renewable energy share to 40% of the European energy production by 2030, compared to the current regulations [2]. Despite the expected market penetration of electricity/hydrogen-powered vehicles and possible energy savings, biofuels and bioliquids will have a significant share in the transportation sector, due to the large long-term demand for dense liquid fuels for aviation, ocean shipping, and long-haul trucking [3]. The expected increase in biofuel production to meet these demands requires assessment of their life cycle impacts, and the implementation of more sustainable biofuel production systems [4].

Bioethanol is a worldwide leader in the transition to more sustainable liquid fuels, used since the 1970s to partially replace gasoline [5]. Its advantages compared to other biofuel alternatives include its cost, global warming potential, accessibility, and socioeconomic impact [6]. An additional advantage is that variable feedstocks can be used for bioethanol production, mainly sugars, starch, and lignocellulosic biomass (e.g., crops, aquatic plants, forest materials, or agricultural residues). The process of converting biomass to ethanol depends on the biochemical composition of this feedstock. The conversion parameters affect the total production costs (including costs for raw material, operation, and capital depreciation) and the cradle-to-grave environmental impacts [7,8].

The extractive fermentation (EF) process combines fermentation with a primary liquid–liquid (LL) extractive separation stage. Removing the produced ethanol and other inhibitory compounds in situ eliminates inhibitions caused by ethanol and increase the ethanol productivity. Recent research acknowledges the potential of this fermentation strategy to reduce operation costs, increase conversion efficiency, and reduce the inhibition of products and by-products. Lemos et al. [9] reported significant productivity increase compared to conventional extraction techniques, demonstrating the potential of EF for use at larger scales. More work is still needed, however, to make EF suitable for industrial application, as shown in recent experimental studies [10,11] on fed batch EF schemes.

Similar to conventional LL extraction processes, EF uses a solvent to selectively extract a desired fermentation product from a mixture. In addition to their extraction efficiency, EF solvents must also be biocompatible with the microorganisms performing the fermentation. When the known solvents (or other tabulated materials) fail to meet the EF process requirements, alternative solvent designs must be sought.

Borhani and Wang [12] classified solvent screening approaches into four different categories according to their reliance on experiments, process and equilibrium models, prediction models, and finally systematic solvent design. Considering the enormous number of candidate materials, trial-and-error using experiments or calculations would be very time-consuming and expensive to apply. Therefore, systematic methods for the optimal selection and molecular design of solvents by means of computer-aided molecular design (CAMD) are becoming increasingly popular for a variety of solvent design applications [13].

Even until recently, contributions in the field of solvent design considered only the solvent properties related to process efficiency. In CAMD, the solvent properties are predicted using quantitative structure-property/activity models (QSPRs/QSARs), to reverse-engineer solvents with desired/optimal properties. For organic solvents, these QSPR/QSAR models usually take the form of functional group-contribution (GC) correlations [14], which provide a good approximation of the property values of unknown materials, using a set of tabulated group-contribution coefficients. Linear GC models are usually applied for [15] the prediction of the solvent physical properties, while UNIFAC models are the dominant choice for accurate prediction of activity coefficients and phase equilibria [12,16].

Given the rising interest in sustainable chemical manufacturing, recent research on CAMD of separation solvents additionally considers health, safety, and environmental aspects [16,17]. For instance, Khor et al. [18] applied a fuzzy optimization approach to the design of palm oil extraction solvents to substitute hexane, considering a set of target properties along with health and safety aspects. Papadopoulos et al. [19] proposed a design framework for phase-change solvents used in CO₂ capture, linked to databases for the evaluation of various sustainability-related indicators. Ten et al. [20] developed a CAMD framework for optimal liquid–liquid extraction solvents considering solvent properties, as well as health and safety criteria using Analytical Hierarchy Process (AHP). Baxevanidis et al. [21] proposed new GC models for lifecycle assessment of environmental impacts and applied them on the CAMD of LL extraction solvents. Solvent optimization studies usually consider a set of solvent properties to improve the process efficiency, e.g., solvent selectivity, solvent capacity, and various primary properties. Detailed process models (if/when applied) are usually considered after CAMD, as a further assessment of the CAMD results [22].

Despite the large progress in the design of separation solvents, challenges remain for the industrial use of the CAMD approach, due to the variety of process-specific properties and the need to capture accurately the desirable solvent profile [17]. This work combines state-of-the art CAMD optimization tools with a rigorous EF process simulator and recently developed lifecycle assessment (LCA) models, to design new, efficient, and sustainable solvents for EF of glucose to bioethanol fuels. The CAMD tool employed here is based on the original implementation of Marcoulaki and Kokossis [23], demonstrated on the design of optimal separation-enhancing organic materials [24]. Since then, the tool has been used extensively on a variety of solvent design applications. These include solvents for reactive separation processes [25], organic Rankine cycle solvents [26], CO₂ capture solvents [27], also considering sustainability aspects [19,21].

Section 2 outlines the CAMD optimizer of Marcoulaki and Kokossis [23]. Section 3 gives the mathematical formulation of the solvent design problem considered here, including the EF process simulator and the GC models required for the prediction of the desirable solvent properties. Section 4 presents solvent design results for different sets of design constraints and objectives. Section 5 discusses the results and Section 6 concludes this work.

2. CAMD Optimization Algorithm

Figure 1 outlines the CAMD optimization algorithm used here. The optimization starts from an initial guess for the functional groups comprising the solvent, e.g., {2xCH3, 3xCH2, 1xCH, 1xOH}. The search proceeds via random perturbations, applied on the initial solvent to generate a new solvent. These perturbations consist of modifying the groups comprising the initial solvent (i.e., addition of new groups, removal of existing groups, replacement of existing groups with new ones). The perturbation strategies and the restoration procedures to ensure structural feasibility of the new solvent are according to Marcoulaki and Kokossis [23].

The new solvent is then compared to the initial one using a set of stochastic criteria according to the Simulated Annealing optimization algorithm [28] as adjusted by Marcoulaki and Kokossis [29]. The comparison is in terms of the performances of the two solvents, based on a set of user-defined objective functions and constraints. For the evaluation of objectives and constraints, the CAMD optimizer is linked to process simulators, GC models, databases, etc., depending on the process task (see Section 3). If the new solvent is accepted, it replaces the initial solvent, otherwise we continue with the initial solvent. This procedure runs iteratively until a set of termination criteria are met, and returns the final solvent, i.e., a final set of functional groups. Note that the typical solution space for this type of problems stretches along trillions of molecular configurations.

Since the mapping between molecules and functional groups is usually many to one, we may have more than one way of connecting the groups together. For example, {2xCH3, 3xCH2, 1xCH, 1xOH} can be any of the following molecules: CCCC(C)CO, CCC(C)CCO, CC(C)CCCO, CCC(CC)CO, CCCCC(O)C, or CCCC(O)CC. Even though the algorithm always generates structurally feasible sets of groups, certain design configurations may be uncommon and potentially hard to synthesize. If this is the case, it is possible to include a complexity assessment in the solvent performance calculations. Additionally or alternatively, one can control the generation of extremely large molecules by limiting the total number of groups, or avoid very complex (possibly unstable) aromatics by limiting the number of rings or the number of substituents per ring, etc. However, setting a very low number of allowable groups may trap the optimizer in a locally optimal domain with limited chance of escaping.

Due to the stochastic nature of the CAMD optimizer, each time we run the tool using the same setup, we may obtain a slightly different final design. The convergence quality of the optimizer is evidenced by the low deviations between the objective function values of the final designs. We also need to ensure that the final objective function value in each run is adequately close (if not the same) with the best objective function value observed during the run.

3. Formulation of the Solvent Design Problem

A good EF solvent has several desirable characteristics. These include biocompatibility, good extraction performance, and low solubility in the aqueous phase, and that the solvent density is adequately different to that of the aqueous phase and that the solvent is chemically stable. Additionally, the solvent selection should contribute to the sustainability of the process, as well as the economic and performance aspects of the process [9]. Based on these criteria, this section presents the mathematical models used here to evaluate the performance of the different solvents generated during the search. The models include a rigorous process simulator to calculate the process objectives, and GC models for the prediction of the various properties appearing in the solvent constraints.

3.1. EF Process Simulator

Several simulation models of variable complexity are available for the mathematical formulation of the EF process, to assist the assessment of candidate extraction solvents [30]. The efficiency of the CAMD tool used here is not affected by the complexity of the process simulator, so we are able to use the rigorous model proposed by Kollerup and Daugulis [31] for continuous stirred tank fermenters. This model is suitable for ethanol production via EF and calculates the extraction efficiency and the productivities in the nutrient and the extract phases.

3.1.1. Simulation Model Assumptions/Requirements

The simulation model assumes the following:

• The broth density, ρ₁, is a linear function of the ethanol and glucose concentration in the aqueous phase. Note that linear mixing rules are the default assumption.
• The bioreaction system produces only cells, ethanol, and carbon dioxide, with yield coefficients YX/S, YP/S, and YC/S, respectively. There are no cells present in the extract phase.
• The specific growth rate, μ, follows the modified Monod equation:

  $$\mu ={\mu }_{\max }\cdot \frac{S_1}{K_S+S_1}\cdot \frac{K_P}{K_P+P_1}$$

  (1)

  where μ_max denotes the maximum specific growth rate, K_P and K_S are the product inhibition and the substrate saturation constants, respectively; S₁ denotes the effluent substrate concentration in the aqueous phase; and P₁ is the effluent product (ethanol) concentration in the aqueous phase.

The simulator model also requires that the solvent is completely biocompatible, and the only function of the solvent with respect to the fermentation system is the extraction of the bioreaction product from the nutrient. It is additionally required that there are negligible solvent losses to the nutrient phase and negligible substrate present in the extract phase. These are solvent-specific requirements enforced under the constraints of Section 3.2.

3.1.2. Simulation Model Formulation

The model inputs are the product distribution coefficient (M_P), the influent substrate concentration in the aqueous phase ($S_1^0$), and the dilution rate in the solvent phase (D_E). The calculation of M_P is discussed in Section 3.2. The $S_1^0$, S₁, and D_E can be fixed, to allow comparison between the candidate solvents under typical fermentation conditions (see Section 4).

The analytical expressions used here in order to calculate P₁ are the following:

$$A\cdot P_1^2+B\cdot P_1+C=0$$

(2)

where the A, B, and C terms are formulated as:

$$A=M_P\cdot \{W_S\cdot \frac{K_P}{{\rho }_P}\cdot \beta -D_E+W_Y\cdot \left(D_E-W_S\cdot \frac{K_P}{{\rho }_P}\right)\}$$

(3)

$$\begin{gathered} B=K_P\cdot \{W_S\cdot \left[\frac{M_P}{{\rho }_P}\cdot \left({\rho }_W+{ \mathrm{S}}_1 \cdot (\alpha -Y_{C/S})\right)-\beta \right]-M_P\cdot D_E+ \\ +W_Y\cdot \left[M_P\cdot D_E-W_S\cdot \left(\frac{M_P\cdot Y_{P/S}\cdot S_1}{{\rho }_P}-1\right)\right]\} \end{gathered}$$

(4)

$$C=W_S\cdot K_P\cdot \left\{\frac{S_1}{S_1^0}\cdot \left[{\rho }_W+S_1^0\cdot \left(\alpha -Y_{C/S}\right)\right]-\left[{\rho }_W+S_1\cdot \left(\alpha -Y_{C/S}\right)\right]\right\}$$

(5)

where

$$W_S={\mu }_{\max }\cdot S_1/(K_S+S_1)$$

(6)

$$W_Y=\frac{{\rho }_W+S_1^0 \cdot \left(\alpha -Y_{C/S}\right)}{Y_{P/S}\cdot S_1^0}$$

(7)

Note that, Equations (2)–(7) are solved for each solvent generated during the search, using available solvers for non-linear equation systems.

3.1.3. Calculation of EF Process Objectives

Having estimated P₁, it is possible to calculate the various process objectives typically used in EF solvent selection, i.e., the ethanol productivities, the substrate conversion, and the extraction efficiency.

The total ethanol productivity is the sum of productivities in the nutrient (PD_W) and the extract (PD_E) phases:

$$P{D}_T=P{D}_W+P{D}_E$$

(8)

The ethanol productivity in the nutrient, PD_W, depends on P₁ as follows:

$$P{D}_W=P_1\cdot D^{\prime }$$

(9)

where the effluent infinite dilution rate in the aqueous phase, D’, is equal to μ of Equation (1).

The ethanol productivity in the extract, PD_E, is given as:

$$P{D}_E=P_2\cdot {D^{\prime }}_E$$

(10)

where P₂ is the product concentration in the extract phase, i.e., $P_2=M_P\cdot P_1$; and D_E’ denotes the effluent dilution rate in the extract phase calculated as:

$${D^{\prime }}_E=D_E\cdot \left[1+{\left({\rho }_P/(M_P\cdot P_1)-1\right)}^{-1}\right]$$

(11)

Using the above equations, the total productivity finally becomes:

$$P{D}_T=P{D}_W+P{D}_E=P_1\cdot \left(D^{\prime }+M_P\cdot {D^{\prime }}_E\right)$$

(12)

The substrate conversion is calculated as follows:

$$C_S=1-\frac{D^{\prime }\cdot S_1}{D\cdot S_1^0}$$

(13)

where D is the influent infinite dilution rate in the aqueous phase, calculated as:

$$D=D^{\prime }\cdot \left(S_1+\frac{P_1}{Y_{P/S}}\right)+\frac{M_P\cdot D_E\cdot P_1}{Y_{P/S}\cdot S_1^0\cdot \left(1-M_P\cdot P_1/{\rho }_P\right)}$$

(14)

The extraction efficiency (EE) is defined as the fraction of PD_E over PD_T, therefore:

$$EE=\frac{D^{\prime }}{D^{\prime }+M_P\cdot {D^{\prime }}_E}$$

(15)

3.2. Constraints for EF Solvents

This section discusses the desirable properties for an appropriate EF solvent, and the models used here for property prediction. Particular emphasis is on the biocompatibility, the product distribution coefficient, and the solvent losses, to satisfy the requirements of the process simulator of Section 3.1.

3.2.1. Distribution Coefficient of the Product to Be Removed (High)

The M_P is the main factor to affect the separation driving force, hence the main criterion in selecting extraction solvents [32]. In the simulator of Section 3.1, M_P participates in the calculations of productivities and EE. According to Pretel et al. [33], the M_P for LL extraction is predicted at infinite dilution conditions and at the expected process operating temperature as follows:

$$M_P^{\infty }=\frac{{\gamma }_{A,B}^{\infty }}{{\gamma }_{A,E}^{\infty }}\cdot \frac{M{W}_B}{M{W}_E}$$

(16)

where ${\gamma }_{i,j}^{\infty }$ denote the activity coefficients of mixture component i in component j; MW_i is the molecular weight of component i; and A, B, and E are the solute, the aqueous feed, and the solvent/extract, respectively. Note that the actual M_P value is taken here as a fixed fraction, η, of $M_P^{\infty }$. Higher accuracy in M_P could be achieved, but this would not alter the relative ranking of the obtained solvent designs.

The most common GC models for the prediction of γ’s are the UNIFAC model and its modifications [34]. Following the usual practice for solvent design, this work applies the original UNIFAC model using the VLE coefficients. The coefficient values for different pairs of functional groups can be found in Reid et al. [35].

3.2.2. Solvent Losses to the Aqueous Phase (Low)

Low solvent solubility to the aqueous phase leads to low solvent losses, and therefore reduces the required solvent makeup and the cost for downstream treatment of the aqueous effluent. In addition, the EF process simulator of Section 3.1 assumes negligible solvent losses to the aqueous phase. Solvent losses, $S_L^{\infty }$, are predicted at infinite dilution conditions and at the expected process operating conditions as follows [33]:

$$S_L^{\infty }=\frac{1}{{\gamma }_{E,B}^{\infty }}\cdot \frac{M{W}_E}{M{W}_B}$$

(17)

Activity coefficients are predicted using the UNIFAC model as explained above.

3.2.3. Solvent Selectivity (High)

High selectivity of the solvent towards the solute is very important in LL extraction processes, and usually completes and/or consolidates the requirement for high M_P. The solvent selectivity, $S_E^{\infty }$, is predicted at infinite dilution conditions and at the expected process operating conditions as follows [33]:

$$S_E^{\infty }=\frac{{\gamma }_{B,E}^{\infty }}{{\gamma }_{A,E}^{\infty }}\cdot \frac{M{W}_A}{M{W}_B}$$

(18)

Activity coefficients are predicted using the UNIFAC model as explained above.

3.2.4. Distribution Coefficient for Essential Nutrients/Byproducts (Low)

The loss of substrate increases the investment/processing costs, and reduces the process efficiency, while the presence of substrate in the extract stream may complicate the recovery/purification stages following the EF stage. In addition, the EF process simulator (see Section 3.1) assumes negligible substrate concentration in the extract phase. Additionally, we may want to limit the mass transfer of reaction byproducts (e.g., acetic acid) from the nutrient to the extract phase. The distribution coefficient of component S (denoting glucose or a byproduct) is given as:

$$M_S^{\infty }=\frac{{\gamma }_{S,B}^{\infty }}{{\gamma }_{S,E}^{\infty }}\cdot \frac{M{W}_B}{M{W}_E}$$

(19)

Activity coefficients are again predicted using UNIFAC VLE as explained above.

3.2.5. Ease of Solvent—Solute Separation (High)

Solvent and solute recovery from the extract mixture is preferably performed by simple distillation. A substantial difference in the boiling point temperatures of the two components can guarantee the absence of azeotropic formations and the ease of separation. Boiling point temperature predictions are according to the well-established GC model of Constantinou and Gani [36], who provide the model formulation and the table of group-contribution coefficients.

3.2.6. Solvent Biocompatibility (High)

Solvent biocompatibility is the most important criterion in extractive fermentation [9]. Biocompatibility is in terms of the toxicity effects of the solvent on the microorganisms that perform the fermentation, so that their population size remains intact. It can be accounted by means of lethal concentration standards [37], or using the solvent octanol/water partition coefficient [31]. Trevizo and Nirmalakhandan [38] investigated the relation between solubilities and toxicity measures. The model used here is the GC model proposed by Gao et al. [39] for lethal concentrations, who provide the model formulation and the table of group-contribution coefficients.

3.2.7. Solvent Environmental Impact (Low)

The European chemicals strategy for sustainability aims to consider global environmental impacts as early as possible in the material design process [40]. There are many aspects of sustainability that can be considered over the lifecycle of a new chemical. We herein consider three climate change indicators: the Global Warming Potential (GWP), the Cumulative Energy Demand (CED), and the EcoIndicator 99 (EI99). Baxevanidis et al. [21] provide the model formulations for these three LCA indices and the tables of group-contribution coefficients. Additional sustainability criteria can be considered, if they can be formulated using the same GC approach.

3.2.8. Other Physical Properties of the Solvent

Other property requirements include that the material should be liquid, thermally stable, and chemically inert at the process conditions. In addition, a large density difference between the nutrient and extract phases enhances immiscibility. All these are predicted using well-established GC models, which are generally available for a wide range of pure-component properties, such as critical properties, phase transition enthalpies, phase change temperatures, heat capacity, and viscosity. Constantinou and Gani [36] provide the model formulations and the tables of group-contribution coefficients for these predictions. These models could also be used to calculate flowrates, equipment size, conditions, etc., but these are beyond the scope of the present study.

3.2.9. Solvent Cost (Low)

Although the solvent cost is certainly extremely important in the selection of materials for large-scale industrial use, it may depend on many external factors making it difficult to estimate the market price of a new material, and this is outside the scope of this study.

4. EF Solvent Design Case Study

4.1. Design Case Parameters

Table 1. Process parameters for the solvent design case study.

Parameter	Value
α	0.411
β	−0.163
ρw	1000 g/L
YX/S	0.10
YP/S	0.46
YC/S	0.44
μmax	0.45 h−1
Kp	23.0 g/L
Ks	1.0 g/L
Tb,P	351.4 K
TF	298 K
S1	30 g/L
η	80%

gives the parameters for the ethanol-water-glucose system used in the EF simulation model of Section 3.1

Table 2. Bounds on the designed solvent properties.

Property	Feasible Range	Bound	Value	Units
$S_E^{\infty }$	≥SE,min	SE,min	2.0	-
$M_P^{\infty }$	≥MP,min	MP,min	1.0	-
$S_L^{\infty }$	≤Slmax	Slmax	0.005	-
$M_S^{\infty }$	≤MS,max	MS,max	0.1	-
Tb,E	∈[Tb,P + Tmin, 600] or [300, Tb,P − Tmin]	ΔΤmin	30.	K
Tf,E	≤TF − ΔΤmin′	ΔΤmin′	10.	K
ρE	≥ρ1⋅Δρ,max	Δρ,max	125%	-
	≤ρ1⋅Δρ,min	Δρ,min	80%	-
tE	≤tE,max	tE,max	1.5, 2.0, 2.5, 3.0, 3.5, 10.	mol/L
GWP	≤GWPE,max	GWPE,max	8.0, 10.0	g CO2-eq/g
CED	≤CEDE,max	CEDE,max	150, 200	kJ-eq/g
EI99	≤EI99E,max	EI99E,max	0.4, 0.5	-

gives the design constraints used here, according to the desired properties for the EF solvent discussed in Section 3.2. The bounds on $S_L^{\infty }$ and $M_S^{\infty }$ enforce the requirement for negligible solvent losses to the nutrient phase (<0.5%) and low substrate distribution in the extract phase (<10%), respectively. The bound on $S_E^{\infty }$ sets the selectivity above 200%. The $M_P^{\infty }$ lower bound is very relaxed, since $M_P^{}$ is controlled via the objective for maximum ethanol productivity. The bounds on the boiling point temperature, T_b,E, allow recovery of the solvent via simple distillation, and also help to avoid solvents with subambient or extremely high T_b,E. We also require that the solvent is liquid at the process conditions, and that the solvent density, ρ_Ε, is significantly different from the broth density, ρ₁. The two sets of bounds for the LCA indices are taken from Baxevanidis et al. [21]. Finally, we consider a range of t_E,max values (from 1.5 to 10 mol/L) to study the effect of t_E on the final designs.

As mentioned in Section 3.1, we use fixed values for the influent substrate concentration in the aqueous phase, $S_1^0$, and the dilution rate for the solvent phase, D_E, to run the simulator.

Table 3. Operational parameters per solvent design case.

Parameter	Parameter Value per Design Case A1 to A12
$S_1^0$ [g/L]	A1, A2, A3, A4				A5, A6, A7, A8				A9, A10, A11, A12
	150				300				600
DE [h−1]	A1, A5, A9			A2, A6, A10			A3, A7, A11			A4, A8, A12
	0.5			1.0			2.0			4.0

gives the data for the operation of the EF process, assuming 12 design cases, denoted as A₁ to A₁₂. The $S_1^0$ and D_E values used here are within the range investigated by Kollerup and Daugulis [31].

According to the availability of tabulated group-contribution coefficients to predict the solvent properties, the functional groups used here are 57 listed in Baxevanidis et al. [21]. Each solvent design is allowed to have up to 40 functional groups and 5 aromatic rings. These numbers are adequately high to facilitate an extensive screening of solvent designs. The optimizer setup is according to Marcoulaki and Kokossis [23,29].

4.2. Solvent Design Results

The solvent design objective considered here is to maximize the ethanol productivity in the extract phase (Equation (10)). The objective is applied to each one of the 12 design cases of Table 3, using 6 different upper bounds for t_E per design case (see Table 2) to study the effect of the toxicity constraint on the productivity of the EF solvents. Having a sample of 10 runs per combination of t_E,max and design case, we end up with 720 computational experiments, which generated a set of 11 final solvent designs, denoted as D₁ to D₁₁. So, some of these solvents work well over a range of operating conditions.

Table 4. Solvent designs obtained for maximum ethanol productivity.

No	SMILES *	tE	ME	Sl × 103	Tb,E	rE	Tf,E	GWP	CED	EI99
D1	ClCCC(=O)OCC#N	1.69	0.671	10.9	498	956	246	3.23	86.2	0.389
D2	CCC(=O)OCCCCC#N	2.08	1.31	4.87	510	756	245	3.76	99.4	0.433 †
D3	C=CCCCCC#N	2.43	1.63	2.15	462	654	221	4.98	141	0.436 †
D4	C=C(CCl)CC#N	2.44	1.31	5.84	465	846	229	3.36	115	0.444 †
D5	CCCCCC#N	2.48	2.46	6.55	442	646	205	3.68	110	0.433 †
D6	C=CC(Cl)CC#N	2.75	1.27	5.16	461	816	223	5.77	151 †	0.442 †
D7	C(C)(C)CCC#N	2.91	2.45	6.53	434	654	190	4.14	132	0.503 ‡
D8	C(C)(C)(C)C(=O)OCCC#N	2.92	1.24	2.28	513	767	252	5.16	153 †	0.562 ‡
D9	C(C)(C)(C)CC#N	3.03	2.52	7.16	425	677	203	4.94	162 †	0.558 ‡
D10	C(C)(C)C(C=O)CCl	4.51	1.59	8.55	445	861	228	6.71	171 †	0.483 †

Notes: * The results of the CAMD tool may relate to more than one molecular structure, and each SMILES presents one of these alternatives; ^† value between the two LCA bounds; ^‡ value outside the LCA bounds of Table 2.

presents the solvents along with their property values predicted using the models of Section 3. The solvents are ranked according to t_E, and they obey the constraints of Table 2 for $S_E^{\infty }$, $M_P^{\infty }$, Sl^∞, T_b,E, T_f,E, and ρ_E. Regarding the LCA indices, different colors in the last three columns of the table indicate how the solutions conform to the GWP, CED, and EI99 bounds. We observe that t_E ranges between 1.69 and 4.51, and cannot go below 1.69 without other constraints being violated. However, the system could tolerate higher toxicities, since the allowable solvent losses (see Table 2) ensure that the solvent amount present in the nutrient phase is extremely small. In terms of the bounds on LCA indices, only D₁ satisfies all the low LCA bounds. All the designs satisfy the lower GWP bound. Most of the designs satisfy the lower CED bound and the rest satisfy the higher CED bound. EI99 is the harder to satisfy, with the majority of designs lying between the two EI99 bounds, and designs D₇, D₈, and D₉ violating the upper EI99 bound.

As explained in Section 2, the output of the CAMD tool is actually a set of functional groups, e.g., D₂ = {1xCH3, 1xCH2COO, 3xCH2, 1xCH2CN}, so the solvent SMILES shown in Table 4 are reasonable design interpretations using these sets. For instance, D₂ could have the ester group closer to the nitrile group (CCCCCC(=O)OCCCCC#N), but usually the farther apart these groups are, the more stable the molecule might be. The proper consideration of stability or inertness is beyond the scope of solvent screening during early design stages, and should be considered at a later stage of validation through laboratory experiments.

Figure 2 shows the effect of the operational parameters of Table 3 on the optimization results when the objective is to maximize the productivity. Since we perform several runs per design case, we obtain a range of values for the ethanol productivity and the extraction efficiency. In the figure, the 12 design cases are first sorted by decreasing D_E and secondly by decreasing $S_1^0.$ We observe that EE is mainly affected by the D_E value, and we can increase the final EE values by increasing the D_E. Looking at Equations (14) and (15) we see that there is a certain dependence between EE and D_E which could justify this trend. Regarding the productivity, we observe a steady reduction of its maximal value (with the exception of case A8) as we first reduce D_E and secondly $S_1^0$.

Detailed information on the obtained productivities and extraction efficiencies for each design case, as well as the generated solvents, can be found at Supplementary Table S1. We can see that under design case A4 ($S_1^0$ = 150 g/L and D_E = 4/h) we obtain the maximum productivities (~100 gr/L/hr) with designs D₅, D₇, and D₉ of Table 4. The associated efficiencies are also very high (~0.97).

Note that, additional results obtained for maximum EE (Equation (15)) indicate that the EE cannot be improved further compared to results using the productivity objective. In addition, the resulting designs are more diverse and complex in their structures, and the possibility of converging to common design features is extremely low. These designs also include the –CH₂CN group, many of them include two nitrile groups; we also get (inferior) solutions featuring aromatic bonds.

5. Discussion of Results

The obtained solvent designs exhibit different extraction efficiencies, productivities, and conversions according to the initial substrate concentration ($S_1^0$) and the dilution rate (D_E). The solvent ranking is maintained regardless of the particular choice of operating conditions, as indicated in Table S1 of the supplementary material. For the productivity objective, the optimal designs (namely D₉, D₅, and D₇) and their ranking remain the same regardless of the design case. The conversion shows strong dependence on the choice of operation, at least in the region of maximal PD_E. We also observe that the most productive solvents are also the most efficient ones. It is, therefore, safe to say that high productivity in the extract phase is accompanied by high EE (at least near the optimum), without any indication that the reverse holds. In general, the productivity objective (Equation (10)) is found to be a very good choice to address the solvent design goals and generate promising solvents.

The solvent molecules give a clear picture of the structures and group compositions that seem to be promising for each design case. The nitrile group “−CH₂CN” is dominant in the designed solvent molecules, regardless of the choice of the objective function and operation. The designs of Table 4 also include ester groups (D₁, D₂, D₈), chloride groups (D₁, D₄, D₆, D₁₀), and double c-bonds (D₃, D₄, D₆). As an indication of the molecules size, their molecular weights range between 100 and 170 g/mole. These design features differ from conventional solvents (e.g., dodecanol) and their types (e.g., large alcohols).

Some of the designed solvents have similar structure and extraction performance. The structural similarities usually refer to chain isomers, and this explains the proximity in their physical property values. Note that, for more elaborate forms of isomerism (i.e., location, cis–trans, and chiral isomerism) the group-wise representation fails to distinguish between the different configurations. However, the physical behavior differences between such isomeric forms are expected to be small.

Designs D₉, D₇, and D₅ have the highest productivities among the results of the productivity objective, and very high extraction efficiencies. They are chain isomers with the general molecular formula C₄H₉−CH₂CN. This opportunity to generalize, demonstrates the consistency of the results and robustness of the CAMD optimizer. In terms of productivity, design D₉ maintains a clear edge, with D₇ and D₅ following with about 0.8% lower productivity. For instance, according to Table S1, the productivities for D₉, D₇, and D₅ are, for set A₄: 100.8, 100.1, and 100.0 g/L/h, and for sets A₇: 40.8, 40.5, 40.5 g/L/h, respectively. Their efficiencies are around 0.97 and 0.93 for data sets A₄ and A₇, respectively. Among the three designs, D₅ has a simple straight chain structure, it is the less toxic (higher biocompatibility), and it has the best LCA indications with GWP, CED, and EI99 at 3.7, 110, and 0.43, respectively. Therefore, D₅ appears as the promising candidate given the current problem objectives.

On the other hand, design D1 is the only one that satisfies the strict set of constraints for all three LCA indices. This design features its best PD_E and EE under design case A4, evaluated at 63.4 and 0.919, respectively. D₁ could also be considered for further investigation, depending on the importance of LCA performance over the design decisions.

Note that, in practice, the analysis of the results may indicate the need for alternative, modified or additional constraints/objectives to guide the screening process. So, we may need to reconsider, reset, and rerun the design problem several times before obtaining a result worth for further investigation. In addition, these choices depend on the application, and may include additional system information once it becomes available. The CAMD tool used here can easily generate a new series of solvent designs, if the constraints, the process data, and/or the process simulation model are modified.

6. Conclusions

This work applies computer-aided molecular design (CAMD) tools to design new efficient and sustainable solvents for extractive fermentation (EF) of glucose to bioethanol fuels. The EF process combines fermentation with a primary liquid–liquid extraction stage, and has the potential to reduce operation costs, increase conversion efficiency, and reduce the inhibition of products and by-products. Finding appropriate solvents to enhance the extraction efficiency without inhibiting the fermentation is an important step towards industrial application of EF. CAMD technologies can perform the solvent screening task in a way that is systematic, general, and robust.

The solvent design procedure presented here uses a state-of-the-art CAMD optimizer to get useful insights on promising molecular structures, and incentives to perform more detailed assessments before a new material is considered for industrial use. The solvent assessment involves a rigorous simulation model for EF, as well as group-contribution methods for the prediction of activity coefficients and various primary properties, toxicity, and LCA indices.

The solvent design studies presented here use 57 different functional groups as building blocks for a great variety of smaller or larger molecules. Despite the opportunities for complexity, the proposed solvent designs are not particularly complex, and indicate a strong preference to medium-sized molecules and particular functional groups. They are also very efficient in meeting the EF process requirements and maximizing the production of ethanol.

The current set up of the problem seems to favor hydrocarbons with chloride and nitrile groups. In principle, the presence of these groups is a strong indication for environmental hazard. Note that, in the current model for solvent selection the considered lifecycle impacts are limited to carbon emissions. Consideration of additional or different environmental impact categories [41] will generate different results [21]. The problem set up is therefore very important to capture the precise design objectives, and the CAMD optimizer used here is not affected by the number or complexity of these models. This is a very important advantage of the approach proposed here, for fast, efficient, and extended screening of solvent alternatives before going to more detailed investigations, using laboratory experiments and detailed process optimization.