A data engineering framework for on-site end-of-life industrial operations

Abstract

Sustainable initiatives for converting end-of-life (EoL) material flows into feedstocks would make a crucial contribution towards protecting our environment and mitigating the negative impacts of anthropogenic activities. Chemical flow analysis enables decision-makers to identify potential environmental releases and exposure pathways at the EoL stage and, therefore, improves the estimation of chemical exposure. Certain industrial facilities apply on-site pollution abatement operations, thereby constituting nodes of the chemical EoL management chain that can be evaluated and improved to enable greater circularity of materials. This work enhances and extends a recently published EoL data engineering framework by using publicly-available databases, data- driven models, and analytic hierarchy approaches to track chemicals, estimate releases, and potential exposure pathways at on-site industrial pollution management operations. The extended framework develops pollution abatement unit (PAU) technologies and estimates their efficiencies, chemical releases, exposure media, operating expenses, and capital expenditures. Relevant case studies based on the food and pharmaceutical industry sectors illustrate the application of the framework for chemical flow allocation and analysis of a chemical of concern and the benefits of integrating and extending the framework with data-driven and multi-criteria decision-making models. The results show how the enhanced framework designs and evaluates PAU technology systems for managing EoL chemical flows and provides release inventories and pathways for conducting chemical risk evaluation and exposure assessment of potential on-site EoL scenarios.

1. Introduction

Every year, humankind and the environment are exposed to chemical substances. Numerous chemicals may present a risk of injury to health or the environment during production, processing, distribution in commerce, use, or end-of-life (EoL) (European Chemicals Agency, 2019; U.S. Environmental Protection Agency, 2017a). Therefore, regulations exist to track and manage chemicals through their life cycle, evaluating potential impacts, and imposing restrictions if needed (Bodar et al., 2018; National Research Council, 2014). Chemical risk assessment supports the selection of safer-profile chemicals and regulatory decision-making to protect human health and the environment (Bernas, 2013; Whittaker, 2015). However, conducting risk evaluation is a time-consuming and challenging task, especially at the EoL stage, due to the extensive data requirements, data scarcity, problem comprehensiveness, proper reporting, traceability, and epistemic uncertainty to describe the risk (Ragas, 2011). The data gap is especially noteworthy when seeking to prioritize thousands of chemicals based on risk (e.g., the U.S. Toxics Substances Control Act inventory lists about 41,000 commercially-active and non-confidential chemicals (U.S. Environmental Protection Agency, 2017b)).

Using publicly-available information, Hernandez-Betancur et al. (2020e) developed a novel EoL data engineering framework for tracking chemicals, estimating releases, and identifying potential exposure scenarios in EoL flows generated at industrial processes and handled at off-site facilities located across the U.S. However, becase of the use of facility-level information for calculating the emission factors, the framework may not allocate releases considering the underlying characteristics of on-site EoL activities and pollution abatement units (PAUs). For instance, the efficiency of a PAU for recovering a chemical of concern from an EoL stream can be essential to allocate chemical flows and exposure downstream of the PAU.

In the literature, it is possible to find methodologies that can be incorporated into the EoL data engineering framework to enhance chemical flow allocation. For example, engineering process design and modeling have represented a toolkit to provide data for life cycle inventory and integrate environmental, health, and safety considerations (Righi et al., 2018; Sugiyama et al., 2008). These approaches can describe a wide variety of PAUs such as boilers, aerobic/anaerobic digesters, wet/dry scrubbers, pH neutralizers, distillers, absorbers, adsorbers, strippers, and biofilters (Baquerizo et al., 2007; Bojarski et al., 2008; Eberle et al., 2017; Jiménez-González et al., 2000; Zhang and Guo, 2013). Likewise, Process System Engineering offers an approach for dealing with complex systems such as solvent recovery and air pollution treatment processes (Cavanagh et al., 2014; Chea et al., 2019; Guerras and Martín, 2019a). Although the above methodologies provide detailed information on the physical, chemical, and biological phenomena underlying each PAU, they require case-by-case development, hands-on application of engineering knowledge, and extensive calculations that can result in a disadvantage compared to the rapid screening provided by the EoL data engineering framework (Smith et al., 2019).

In contrast, data-driven modeling can help to streamline the chemical flow allocation and to overcome the case-by-case limitation (Cashman et al., 2016; Meyer et al., 2019). For example, data mining has proved to reduce the cost associated with screening life cycle assessment (Sundaravaradan et al., 2011), while machine learning addresses data deficiencies (Song, 2019; Zhu et al., 2020). The data-driven method relies on the company data, which can offer some benefits in terms of more realistic insights (Li et al., 2018). However, it can present limitations because of flow allocation and coverage of reported substances, which lead to data gaps (Smith et al., 2017). The performance of these models depends on the specific dataset domain. Diverse sources can supply the datasets for data-driven modeling. Although some databases like EXIOBASE are free to use, others are licensed and proprietary (Greendelta, 2013), e.g., ecoinvent, which is widely used worldwide for life cycle inventory and impact assessment (Vélez-Henao et al., 2020). Nevertheless, their methodologies focus on relating products and services to cradle-to-grave environmental impacts instead of chemical flow tracking (Frischknecht et al., 2007).

This contribution describes the addition of new data sources, data- driven and multi-criteria decision-making (MCDM) models to enhance and extend the EoL data engineering framework to track chemicals, estimate releases, and potential exposure pathways at facility on-site PAUs. This work has four key features to provide relevant data and PAU techno-economic information into the framework. First, collecting information regarding PAU technologies such as the removal efficiency related to substances of interest and the predominant phase of EoL flow managed by the PAU (e.g., liquid waste). Second, leveraging the data itself and literature for filling data gaps and allocating chemical flows. Third, transforming the information into a machine-readable structure for PAU design, evaluation, and selection using data-driven and MCDM models. Fourth, incorporating the created features into the framework developed by Hernandez-Betancur et al. (2020e) to include inside-facility allocation, new data sources, and tools for decision-making. The enhanced framework can offer insights for selecting and recommending PAU technologies to handle chemicals of concern in EoL streams at the early engineering design step to achieve a sustainable process and a circular life cycle (Hassim, 2016). Furthermore, the enhanced framework may be used as a first layer to provide alternatives for building optimization approaches for pollution abatement systems and networks to deal with chemical pollutants (Chea et al., 2020; Guerras and Martín, 2019b; Yenkie et al., 2019).

2. Methodology

For a chemical of concern c in an EoL input flow to a PAU (F_input), the framework allocates the chemical into the waste/environmental release flow ($F_{\mathit{\text{waste/release}}}^{(c)}$), the remaining chemical flow leaving the PAU, i.e., effluent ($F_{\mathit{\text{effluent}}}^{(c)}$), and determines the chemical flow being destroyed ($F_{\mathit{\text{destroyed}}}^{(c)}$, including converted and degraded), removed ($F_{\mathit{\text{removed}}}^{(c)}$), or recycled ($F_{\mathit{\text{recycled}}}^{(c)}$) as well as the chemical fugitive air release ($F_{\mathit{\text{fugitive}}}^{(c)}$). This allocation considers the PAU effect on the chemical like thermal (Lee et al., 1986; National Research Council, 2000; Saxena and Jotshi, 1996), chemical (Fanning, 2000; Huang et al., 1993; Wang et al., 2005), biological (Darvin and Serageldin, 2003; Kaur, 2017; Tay and Zhang, 1999), and physical (Jaeger Products Inc, 2010; Kreith et al., 2001; U.S. Environmental Protection Agency, 2002) treatments and energy (Energy and Environmtental Analysis Inc., 2005; Mantus, 1992; Ottoboni et al., 1998) and material recovery processes (Bascone et al., 2016; Chmielewski et al., 1997; Hansen et al., 1991; Mular et al., 2002; Shin et al., 2009; Smallwood, 2002; U.S. Environmental Protection Agency, 1978). Additionally, the framework incorporates the predominant phase for $F_{\mathit{\text{input}}}\left({\alpha }_{\mathit{\text{input}}}\right)$ because is a key PAU technology selection criterion.

As presented in Fig. 1, the framework employs an input-output model representation for each PAU. This model must satisfy the annual material balance for a chemical of concern c in Equation (1).

$$\begin{gathered} F_{\mathit{\text{input}}}^{(c)}={\eta }^{(c)}\times F_{\mathit{\text{input}}}^{(c)}+F_{\mathit{\text{fugitive}}}^{(c)}+F_{\mathit{\text{waste/release}}}^{(c)}+F_{\mathit{\text{effluent}}}^{(c)} \\ {F}_{\mathit{\text{input}}}^{(c)}=w^{(c)}\times F_{\mathit{\text{input}}} \\ {F}_{\mathit{\text{fugitive}}}^{(c)}=\left(1-{\eta }^{(c)}\right)\times {\beta }^{(c)}\times F_{\mathit{\text{input}}}^{(c)} \\ {\eta }^{(c)}\times F_{\mathit{\text{input}}}^{(c)}=\{\begin{array}{ll}{F}_{\mathit{\text{recycled}}}^{(c)},\hfill & \mathit{\text{if recycling}}\hfill \\ F_{\mathit{\text{removed}}}^{(c)},\hfill & \mathit{\text{if physical treatment}}\hfill \\ F_{\mathit{\text{destroyed}}}^{(c)},\hfill & \mathit{\text{otherwise}}\hfill \end{array} \end{gathered}$$ (1)

$w^{(c)}$ represents the input concentration of the chemical in F_input, $F_{\mathit{\text{input}}}^{(c)}$ the chemical input flow, ${\beta }^{(c)}$ the emission factor for $F_{\mathit{\text{fugitive}}}^{(\text{c})}$, and ${\eta }^{(c)}$ the PAU abatement efficiency relative to the chemical, i.e., the degree to which the PAU destroys, removes, or recovers it. The framework tracks the chemical regardless of other incidentals that might be generated during the EoL activity, as the case for combustion, chemical, and biological operations that may generate undesired substances. ${\beta }^{(c)}$ is calculated as described in Hernandez-Betancur et al. (2020e), which uses facility-level information for estimating the environmental chemical releases.

Fig. 1.

Generic PAU diagram for tracking chemical flows based on the PAU effect on a chemical of concern. The framework allocates output flows into $F_{\mathit{\text{waste/release}}}^{(c)}$, $F_{\mathit{\text{effluent}}}^{(c)}$, $F_{\mathit{\text{fugitive}}}^{(c)}$, and $F_{\mathit{\text{destroyed}}}^{(c)}$, $F_{\mathit{\text{removed}}}^{(c)}$, or $F_{\mathit{\text{recycled}}}^{(c)}$.

The framework focuses on recycling, energy recovery, and treatment activities. Also, the framework enables exchanging and integrating information from siloed publicly-available databases and transforms information into a machine-readable structure for future automatization, thereby creating the PAU dataset. A GitHub Repository named PAU4Chem has the Python scripts that build the PAU dataset (Hernandez--Betancur et al., 2020a). The PAU dataset information allows input-output modeling to track and allocate chemicals of concern. Moreover, the PAU dataset has chemical unit prices (UP^(c)) and EoL activity capital expenditures (CAPEX) and operating expenses (OPEX). Hence, the generic framework can provide more realistic insights into the relationship between PAU technologies, expenses, and chemical allocation.

In order to develop the PAU dataset, the framework integrates the following databases: Toxics Release Inventory (TRI) (U.S. Environmental Protection Agency, 2013a), Chemical Data Reporting (CDR) (U. S. Environmental Protection Agency, 2016), Facility Registry Service (CDR) (U.S. Environmental Protection Agency, 2004), Statistics of U.S. Businesses (SUSB) (U.S. Census Bureau, 2011, 2007), Annual Survey of Manufactures (ASM) (U.S. Census Bureau, 2009), and Pollution Abatement Costs and Expenditures (PACE)–Survey (U.S. Environmental Protection Agency, 2017c). Additionally, it uses the North American Industry Classification System (NAICS) structure to cluster the TRI reporting facility into industry sectors (ISs) to connect the TRI information to the SUSB, ASM, and PACE Survey (U.S. Census Bureau, 2017a). Three modules constitute the PAU dataset: Technologies, OPEX & CAPEX, and ChemPrices. Sections 2.1, 2.2, and 2.3 depict how the framework effectively gathers, cleans, transforms, and integrates qualitative and quantitative information from multiple data sources to build the three modules. Section 2.4 shows how to use the PAU dataset modules for data-driven modeling and incorporating MCDM to predict EoL management and chemical flow analysis (CFA).

2.1. Data engineering for collecting PAU information: Technology module

Fig. 2 illustrates the steps for building the PAU dataset – Technology module. The framework harnesses the TRI Program (from 1987 to 2018) as this module backbone due to its comprehension and data availability and granularity that enables gathering information for the PAU technologies (U.S. Environmental Protection Agency, 2014).

Fig. 2.

Data engineering for transforming information from TRI database into structures for the PAU dataset – Technologies and CFA.

In step one, PAU technologies and uses/activities for a given chemical c are collected. For treatments, PAU sequences, codes describing α_input, $w^{(c)}$ and ${\eta }^{(c)}$ value ranges, and ${\eta }^{(c)}$ estimated values are taken. However, $w^{(c)}$ and ${\eta }^{(c)}$ estimated values were reported from 1987 to 2004 (U.S. Environmental Protection Agency, 2013b). If treatment activities for the chemical c were submitted by a facility, typical $w^{(c)}$ and α_input are estimated for recycling and energy recovery activities. Otherwise, as NAICS uses a hierarchical structure organized from 2-digit NAICS codes (less specific) to 6-digit NAICS codes (more specific) (U. S. Census Bureau, 2017b), this structure supports estimating $w^{(c)}$ and α_input. Moreover, higher weight is given for typical α_input. Records with no estimated $w^{(c)}$ and α_input are dropped (step two).

In step three, ${\eta }^{(c)}$ is estimated for recycling and energy recovery activities. Due to energy recovery and incineration activities are combustion operations, incineration activities from TRI are leveraged to obtain a ${\eta }^{(c)}$ value. Using the NAICS structure as in step one, ${\eta }^{(c)}$ is assigned. However, if ${\eta }^{(c)}$ cannot be estimated, this is obtained by considering the threshold limits under hazardous waste and hazardous air pollutant regulations (U.S. Environmental Protection Agency, 2006, 1999).

For recycling activities, the EoL chemical flows reported by each facility are used to estimate a ${\eta }^{(c)}$ potential value. As depicted in Fig. 3, an EoL chemical flow may have several potential pathways like transfers to sewage treatment plants. The exact pathway a chemical follows in a facility is uncertain. To overcome this uncertainty, for each facility, the non-zero flows for a chemical c are combined obtaining a set of scenarios Θ. Each scenario θ_i ∈ Θ has a possible input flow to recycling $F_{\mathit{\text{input}},{\theta }_i}^{(c)}$. Using $F_{\mathit{\text{input}},{\theta }_i}^{(c)}$ and the reported $F_{\mathit{\text{recycled}}}^{(c)}$, ${\eta }_{{\theta }_i}^{(c)}$ is calculated. Assuming a recycling operation is used to obtain high cost-effectiveness, Equation (2) takes the upper value of ${\eta }_{{\theta }_i}^{(c)}$ as ${\eta }^{(c)}$, if this value is not an outlier and its coefficient of variation (CV) is less than 1 (low variance), ensuring a narrow value range for ${\eta }^{(c)}$. If the above two conditions are not satisfied for a facility, a ${\eta }^{(c)}$ value is assigned leveraging the NAICS structure again. The assignment considers facilities having the same recycling PAU technology type.

Fig. 3.

Schematic explanation to calculate ${\mathit{\eta }}^{(c)}$ for recycling using TRI information.

$${\eta }^{(c)}{=}_{{\theta }_i\in \text{Θ}}^{MAX}\left({\eta }_{{\theta }_i}^{(c)}\right){=}_{{\theta }_i\in \text{Θ}}^{MAX}\left(\frac{F_{\mathit{\text{recycled}}}^{(c)}}{F_{\mathit{\text{input}},{\theta }_i}^{(c)}}\times 100\right)$$ (2)

In step four, the framework excludes the PAU dataset records if ${\eta }^{(c)}$ could not be estimated. Finally, in step five, the value range for F_input is calculated using the $w^{(c)}$ range, ${\eta }^{(c)}$, and $F_{\mathit{\text{recycled}}}^{(c)}$. Hence, this module provides F_input, α_input, $w^{(c)}$, ${\eta }^{(c)}$, PAU technologies, and chemical uses/activities.

2.2. Data engineering for estimating OPEX and CAPEX: OPEX & CAPEX module

OPEX and CAPEX influence implementing a PAU technology (Collins and Harris, 2002; Gray and Shadbegian, 1995, 1998). Hence, incorporating both parameters would result in realistic insights for inferring/predicting whether an IS affords a specific PAU technology. The framework uses the 2005 PACE–Survey to obtain the CAPEX and OPEX (U.S. Environmental Protection Agency, 2017c). The PACE–survey presents such parameters by media (air, water, and solid waste), EoL activity, and IS (only manufacturers). However, using the publicly-available version is impossible to know each surveyed facility expenses in USD/EoL-flow-kg, e.g., treating air emissions since the information is presented as an aggregated value in USD. Fig. 4 shows a procedure based on Monte Carlo for mimicking potential CAPEX and OPEX in USD/EoL-flow-kg for each facility to manage contaminants that would otherwise have polluted the environment.

Fig. 4.

Data engineering for mimicking the PACE–survey and obtaining the PAU dataset–OPEX & CAPEX module.

In step one, the probability of sampling an IS is calculated. The PACE–Survey targeted ISs with a high percentage of facilities reporting no OPEX in 1994. Hence, the framework selects 1994 TRI and uses the EoL chemical flows to determine potential facilities in ISs with no OPEX in 1994. In step two, the probability of sampling a facility is estimated, considering that the 2004 SUSB was the PACE–Survey sample frame and facilities with twenty or more employees were eligible.

The CAPEX and OPEX correlate positively with the value of shipments (VoS) (U.S. Environmental Protection Agency, 2017c), which measures the USD of products sold by manufacturers, and it is an IS measure-of-size. Thus, the probability of sampling a facility from an IS is proportional to the VoS. The 2008 SUSB and 2008 ASM were selected due to being the closest survey period to 2005. The total VoS, including relative standard error, is taken from the 2008 ASM, meanwhile, the number of facilities from the 2008 SUSB. Using these two values, the VoS mean and standard deviation by IS are calculated. After using the sample frame and the above VoS statistical measures, a lognormal distribution is obtained to assign a VoS for a facility. This distribution has positive values that create a right-skewed curve explaining better the earnings behavior (Heckman and Sattinger, 2015).

In step three, facilities are drawn from the sampling pool, considering the probabilities calculated in steps one and two. In step four, from the PACE–Survey, the CAPEX and OPEX by activity and media are used to calculate the probability that facilities within ISs may have spent on specific activity and media. Thus, using the above probability, the one calculated in step two, and assuming these two events are independent, the number of facilities within ISs having at least a PAU for a specific activity and media are determined.

Finally, in step five, the 2004 PAU dataset–Technologies helps calculate the EoL flow mean and standard deviation by IS, media, and activity to obtain lognormal distribution to assign an EoL flow value for a facility. EoL flow by IS, activity, and media is calculated by the probability obtained in step four and summing up the EoL flow values. This amount normalizes the OPEX and CAPEX for each activity, media, and IS, i.e., OPEX and CAPEX in USD/EoL-flow-kg.

2.3. Data engineering for estimating chemical unit price: ChemPrices module

ChemPrices module contains information about the relationship between the PAUs and the chemical/chemical category prices (UP^(c)) in USD/g. UP^(c) is obtained from e-commerce sources like SciFinder, Amazon, Alibaba, and Fisher Scientific, considering the currency exchange rate of the U.S. dollar. Due to having UP^(c) for multiple suppliers, the framework drops the outliers using the Z score test. For the TRI chemical categories, the framework uses regulatory lists to know potentially-candidate chemicals belonging to them. Such lists can be found in the GitHub repository PAU4Chem (Hernandez-Betancur et al., 2020a). The framework uses the FRS, a database containing information about facilities regulated by the U.S. Environmental Protection Agency. FRS provides the alternative name for the reporting facilities in the PAU dataset so the framework can connect them to the CDR database. Hence, the framework identifies those chemicals reported by the facilities to the CDR and allocates these into the TRI chemical categories. If the search is successful, the framework calculates UP^(c) for a category using a smaller group of chemicals; otherwise, it uses all the chemicals in the corresponding category. UP^(c) attribution is based on the median because it is a central tendency statistic less sensitive to outliers than the mean.

2.4. Estimation of EoL management and CFA

This section shows a procedure to employ the PAU dataset modules with data-driven and MCDM to suggest PAU technologies and estimate their ${\eta }^{(c)}$, CAPEX, and OPEX. Moreover, the procedure moves forward to perform CFA after the MCDM, thus, enabling the rapid estimation of chemical releases and output streams from the PAU technologies and providing exposure scenarios for further assessment (Hernandez-Betancur and Ruiz-Mercado, 2019).

Fig. 5 shows the step-by-step procedure to perform CFA for a chemical of concern in PAU sequences. Bayesian Networks (BNs) infer the potential PAU technologies to manage the chemical (Koller and Friedman, 2009). BN variables are the CAPEX, OPEX, type of EoL management, ${\eta }^{(c)}$, $w^{(c)}$, F_input, UP^(c), α_input, whether the chemical is an impurity/by-product ($I_{IB}^{(c)}$ = Yes/Not), and PAU technology. The BNs structure, variables, and their connections are in the GitHub repository (Hernandez-Betancur et al., 2020b). This structure mimics the decision-making process of technical stakeholders selecting and designing a PAU. The three PAU dataset modules supply the data for building the conditional probability tables for revealing relationships between the BN variables, e.g., the conditional probability of obtaining a ${\eta }^{(c)}$ value given $w^{(c)}$. As shown in Fig. 5, if F_input has n individual chemicals, n BNs are built. Thus, the inference is based on a case-by-case analysis using the individual chemical information from the PAU dataset. The BNs only use the PAU dataset from 1987 to 2004 since these reporting years have information about $w^{(c)}$. Moreover, the procedure uses the PAU dataset to determine the PAU sequence using only dataset records not reporting PAU sequences to avoid any systematic error. Stakeholders must enter either ${\eta }^{(c)}$ or PAU technology or optional problem specifications to calculate the probability of either selecting a PAU technology or ${\eta }^{(c)}$ for a chemical.

Fig. 5.

A generic step-by-step procedure and schematic explanation to perform CFA in PAU sequences by using the PAU dataset and data-driven and MCDM models.

As Fig. 5 depicts, if a chemical has more than one potential PAU technology to satisfy the specifications, the Fuzzy Analytic Hierarchy Process (FAHP) with triangular fuzzy numbers supports the MCDM (Hernández-Betancur et al., 2019). The FAHP–selection uses the criteria presented in Fig. 5, considering the importance of the type of EoL management activity to which the PAU technology belongs, i.e., if the PAU technology is for recycling (more preferable), energy recovery, or treatment (less preferable) (U.S. Environmental Protection Agency, 2015). The FAHP–selection checks the probability of a type of EoL management, and the PAU technology are selected and designed by the BN. If the F_input has more than one chemical, the FAHP–selection considers whether the PAU technology can manage several chemicals in F_input. Finally, if CAPEX and OPEX are not problem specifications, they are considered as FAHP-selection criteria.

As mentioned before, the framework considers the effect of the PAU technology on the chemical of concern. Hence, if more than one PAU technology is needed to manage an EoL stream, the FAHP–sequence supports PAU sequences arrangement. FAHP–sequence works using five criteria. Three criteria are chemical flammability, instability, and corrosiveness. These criteria ensure the safety of equipment and process structure. These properties are in a public GitHub repository (Hernandez-Betancur et al., 2020c), supporting exposure assessment and circular life cycle endeavors (Hernandez-Betancur et al., 2020d). The fourth criterion is the feasibility of finding a similar PAU technology sequence in the PAU dataset. The fifth criterion is $F_{\mathit{\text{input}}}^{(c)}$, associated with PAU equipment size and cost.

Finally, the CFA is completed considering Equation (1) and the procedure for calculating ${\beta }^{(c)}$ developed by Hernandez-Betancur et al. (2020e). As described in Fig. 5, the CFA complies with boiling and melting points to determine whether a chemical is assessed as a liquid, solid, or gas at standard conditions. The CFA considers if the chemical is a metal to determine whether combustion operations may abate it. Based on PAU technology functioning knowledge, α_input, and water solubility, a potential predominant phase for the output streams (α_output) is assigned. Meanwhile, the physicochemical properties are in the GitHub repository named above (Hernandez-Betancur et al., 2020c). The procedure and data-driven and MCDM depicted in Fig. 5 are not definitive. In the future, it would be extended and modified to predict potential on-site EoL management for chemicals that are not in the PAU dataset.

3. Case studies

Three relevant case studies based on food and pharmaceutical ISs illustrate the application of the enhanced and extended framework with the PAU dataset modules described in Sections 2.1 to 2.3 and the procedure presented in Section 2.4. The results show how the enhanced framework designs and evaluates PAU technology systems handling individual and multiple chemicals in EoL flows and provides release inventories and pathways for conducting chemical exposure assessment for potential on-site EoL scenarios. The case studies specify budget (CAPEX and OPEX), ${\eta }^{(c)}$, $w^{(c)}$, F_input, UP^(c), α_input, and $I_{IB}^{(c)}$. Case study Python scripts can be found in a public GitHub repository (Hernandez-Betancur et al., 2020b). Tables describing relevant input/output information for the case studies are in the supporting information.

The chemicals for the case studies are isopropanol, methanol, ammonia, ethylene glycol, n-hexane, toluene, N,N-dimethylformamide, and dichloromethane. These eight chemicals are part of the Organization for Economic Co-operation and Development List of High Production Volume Chemicals, which contains chemicals produced in amounts equal to or greater than 1,000,000 kg/yr (Organization for Economic Co-operation and Development, 2004). Those chemicals can be found at least in one of the lists of chemicals of concern categories, e.g., as hazardous air pollutants, hazardous wastes, and extremely hazardous substances (U.S. Environmental Protection Agency, 2020). Chemicals like methanol, ammonia, ethylene glycol, and n-hexane are associated with toxic releases from the food manufacturing IS (Gaona et al., 2020). Isopropanol, methanol, toluene, N,N-dimethylformamide, and dichloromethane are related to pharmaceutical manufacturing. These four chemicals are widely released into the air and incinerated (Beck et al., 1978).

Hence, as outcomes from in-depth numerical validation findings, the case study results show features, implications, and limitations of the data engineering framework that should be considered in future developments. Case study 1 demonstrates that, although using historical data can increase data availability, it can dilute most recent data evidencing releases reduction due to using more cleaner technologies. Case study 2 shows potential unintended outcomes on the predictions due to using ranges to describe $w^{(c)}$. Case study 3 shows that the framework can provide data to build models capable of designing PAU technologies and arranging PAU sequences like those used in actual industrial facilities.

4. Results and discussion

4.1. Case study 1: Celecoxib manufacturing process

Three EoL flows associated with the isopropanol/water washes, mother liquor or filtrate, and dryer distillates from the Celecoxib manufacturing process are used as a case study (Hounsell et al., 2012; Slater et al., 2008). These streams contain water, methanol, ethanol, and isopropanol. Isopropanol and methanol are the chemicals of concern selected due to their toxicity and presence in the PAU dataset. The GitHub repository has all the problem specifications for the BNs, the criteria values for performing the FAHP decision-making, and the physicochemical properties for completing the CFA (Hernandez-Betancur et al., 2020b). The input chemical concentration ($w^{(c)}$) and the predominant input phase (α_input) may determine the type of management for each chemical in the EoL flows. Table S1 presents these specifications. The EoL input flow to a PAU (F_input) is 8.05 × 10⁶ kg/yr. This value is the average F_input containing both chemicals and is calculated using the estimated values of F_input for the pharmaceutical industry sector (IS) records from the PAU dataset.

A finding from the development of the case studies relates to the best procedure for performing the CFA for the PAU/PAU sequence by following Equation (1). A PAU-level approach is employed to perform the CFA and ensures that both the material balance and the required abatement efficiency relative to the chemical (${\eta }^{(c)}$) are satisfied (bottom- up in Table S2). This is an enhancement to the framework developed by Hernandez-Betancur et al. (2020e) since using facility-level information to estimate emission factors (top-down in Table S2) causes an overestimation of $F_{\mathit{\text{fugitive}}}^{(c)}$ and does not meet the ${\eta }^{(c)}$ requirements.

Comparing between the CFA obtained by both approaches for stream # 1 in Table S1, e.g., the mean $F_{\mathit{\text{fugitive}}}^{(c)}$ for isopropanol from the batch still distillation is 1.99 × 10⁴ kg/yr using the bottom-up approach. In contrast, $F_{\mathit{\text{fugitive}}}^{(c)}$ is 9.86 × 10⁵ kg/yr using the top-down approach. The mean $F_{\mathit{\text{destroyed}}}^{(c)}$ for isopropanol is 3.93 × 10⁶ kg/yr by the bottom-up approach, which means 99.18% of the isopropanol fed in stream # 1. Instead, for the case of the top-down approach, $F_{\mathit{\text{destroyed}}}^{(c)}$ is 1.97 × 10⁶ kg/ yr, i.e., 49.86% of the isopropanol fed into stream # 1 (see comparison Table S2 for more details). In this case study, the expected ${\eta }^{(c)}$ for isopropanol must be 99.50%; therefore, the bottom-up approach can satisfy the ${\eta }^{(c)}$ established in the material balance. Hence, this exercise demonstrates the effectiveness of using the bottom-up approach to perform all material balances and design PAU systems to estimate chemical releases at on-site EoL management activities, as proposed in this research contribution. Therefore, the inclusion of the on-site EoL activities and PAU technology information enhances the CFA at the EoL stage developed by Hernandez-Betancur et al. (2020e).

Considering the problem specifications like $w^{(c)}$, α_input, and ${\eta }^{(c)}$ for the three streams from the Celecoxib process (see Table S1), the framework suggests batch still distillation for recycling methanol, while incineration using liquid injection for treating isopropanol, see further details in Table S3. Hence, the PAU dataset used to develop the BNs would lead to suggest to a stakeholder the use of a technology associated with a destructive process like a liquid injection incinerator. In fact, from a phenomenological perspective, liquid injection incinerator is a suitable technology for wastes with a high-organic content like the three streams from the Celecoxib process (Lee et al., 1986). The BN has a structure whose estimates are based on statistical evidence. Hence, the statistical evidence in the PAU dataset from 1987 to 2004 shows the companies used to destroy the isopropanol instead of recovering it. In fact, researchers have studied isopropanol recovery by pervaporation and distillation as an alternative to incineration (Hounsell et al., 2012; Slater et al., 2008). Methanol has been reported several times, describing successful applications of green chemistry and engineering activities (U. S. Environmental Protection Agency, 2019).

Although both methanol and isopropanol are on the Organization for Economic Co-operation and Development List of High Production Volume Chemicals, the quantities of methanol reported in the U.S. as domestically manufactured, imported, used, and exported far exceed those of isopropanol. For example, the ratio of imported methanol to imported isopropanol is about 51.73 (see Table S9). This may explain why the statistical distribution of the data used to construct the BN leads to the selection of recycling for methanol but not for isopropanol. However, the quantities may also reflect the importance of the methanol market relative to that of isopropanol. This may be supported by the fact that the global market size of the methanol is around USD 33.69 billion, while for isopropanol it is around USD 2.65 billion (see Table S9). Nevertheless, unlike BNs, other data-driven models combined with a data preprocessing step for the PAU dataset can get a greener suggestion for managing EoL flow containing isopropanol or even other substances, thereby overcoming the data drift and concept drift that could affect the applicability of the models.

An important implication from a life cycle inventory perspective and chemical releases quantification is that after selecting PAU technologies, the framework suggests potential output streams and their α_output where the chemical in the case study may be allocated downstream of the PAUs. This implication is important for understanding the potential indirect risk that may be associated with using a PAU technological to abate a chemical of concern. For example, methanol may generate $F_{\mathit{\text{waste/release}}}^{(c)}$ (whose α_output is wastewater) from batch still distillation. Both methanol and isopropanol may be in $F_{\mathit{\text{effluent}}}^{(c)}$ from the liquid injection operation related to stack air releases. In addition to the potential output streams and α_output, the framework also provides a value range for each flow (see Table S3). For instance, the mean $F_{\mathit{\text{fugitive}}}^{(c)}$ for isopropanol from the batch still distillation is 1.99 × 10⁴ kg/yr for the stream # 1 and 1.39 × 10⁴ kg/yr for stream # 2. The result above means the $F_{\mathit{\text{fugitive}}}^{(c)}$ for isopropanol from batch still distillation is higher for stream # 1 than for stream # 2. This result makes sense, considering $w^{(c)}$ for isopropanol is higher for stream # 1, and F_input for both cases is 8.05 × 10⁶ kg/yr. Moreover, the mean value is accompanied by a Coefficient of Variation (CV) for each flow estimate (see Table S3), enabling the analysis of intervals instead of point values for the results. For example, methanol in $F_{\mathit{\text{waste/release}}}^{(c)}$ from batch still distillation can be in the interval of [0, 9.49 × 10⁰ kg/yr] and [0, 1.31 × 10¹ kg/yr] for stream # 1 and 3. Hence, it may be possible to find methanol quantities in $F_{\mathit{\text{waste/release}}}^{(c)}$ from batch still distillation for stream # 1 lower than for stream # 3, even though $w^{(c)}$ for methanol in stream # 1 is higher than for stream # 3, as shown in Table S1.

4.2. Case study 2: Solvent EoL flows from Food IS

Methanol, ammonia, ethylene glycol, and n-hexane are four chemicals widely used in the food IS (Gaona et al., 2020). Case studies have been addressed using possible values for the problem specifications (Hernandez-Betancur et al., 2020b). Information for these four chemicals reported by this IS was searched in the PAU dataset to identify whether the chemical is an impurity/by-product ($I_{IB}^{(c)}=Yes$), $w^{(c)}$, and the average F_input. Hence, the F_input containing the chemicals is fixed to 8.28 × 10⁶ kg/yr for methanol, 2.60 × 10⁹ kg/yr for ammonia, 8.70 × 10⁹ kg/yr for ethylene glycol, and 5.40 × 10⁸ kg/yr for n-hexane. Each F_input is assumed to be a methanol-water mixture (see Table S4). The desired output chemical concentration of the case study is based on the lowest ecological benchmark found for each chemical in the Risk Assessment Information System (U.S. Department of Energy, 1998), except for methanol that is based on a study developed for the American Methanol Institute (Malcolm Pirnie Inc, 1999). These values are used to set the required ${\eta }^{(c)}$ for each circumstance. Hence, setting parameter $I_{IB}^{(c)}$, $w^{(c)}$, ${\eta }^{(c)}$, and α_output to a specific value could lead to suggest a different PAU technology for each F_input (see Table S4 for the values established).

Performing the procedure shown in Fig. 5 results in the selection of PAU technologies for each solvent F_input from the food IS and their CFA for allocating the chemicals downstream (see Tables S5–S8 for more details). An aspect to highlight is that according to the results obtained for methanol, see Table S5, and ammonia, see Table S6, the selection of the PAU technology depends more on the input concentration of the chemical ($w^{(c)}$) and the required PAU abatement efficiency relative to the chemical (${\eta }^{(c)}$) than whether the chemical is an impurity/by-product ($I_{IB}^{(c)}=\mathit{\text{Yes}}$). For instance, batch still distillation is selected for managing a solvent F_input when $w^{(c)}$ for ammonia is 0.01 %wt/wt regardless of $I_{IB}^{(c)}$ = Yes for ammonia. Likewise, solvent recovery via fractionation is selected when $w^{(c)}$ for ammonia is 0.01 and 0.51 %wt/wt. Hence, for data-driven modeling and decision-making, $w^{(c)}$ has higher relative importance than $I_{IB}^{(c)}$.

Another variable having an important influence on data-driven modeling and decision-making is F_input. For example, in case study one, F_input containing methanol was 8.05 × 10⁶ kg/yr, and the result suggested solvent recovery via batch still distillation. Likewise, for streams # 4 and 7, F_input is 8.28 × 10⁶ kg/yr, and the suggestion in both cases is batch still distillation for methanol recovery. In contrast, for streams # 11, 12, 15, and 16 containing ammonia, F_inputis 2.60 × 10⁹ kg/ yr, and the recovery for all the cases would require fractionation. The above may be explained because continuous distillation and fractionation are often preferred to batch operations for large solvent recovery streams (Douglas, 1988; Smallwood, 2002). This aspect may lead to developing future data-driven models that assign a high weight to the parameters associated with $w^{(c)}$ and F_input; however, this aspect should be studied thoroughly.

The PAU selection in case of the highest $w^{(c)}\text{s}$ indicates a high chance of selecting treatment via incineration when $w^{(c)}$ is high. For instance, Table S5 indicates when $w^{(c)}$ for methanol is 75.00 %wt/wt, liquid injection incineration operation is chosen, which is like what happens with n-hexane at the same $w^{(c)}$, as indicated in Table S8. In case study one, the same occurred with isopropanol, whose $w^{(c)}$ was between 34.50 and 50.70 %wt/wt, close to high organic content, the desired requirement for combustion operations. Although the above F_input contains water, liquid injection incineration can handle high organic-strength aqueous wastes without requiring auxiliary devices (Lee et al., 1986). As explained in case study one, facilities should explore recycling and reuse activities instead of using incineration to reduce the waste volume before disposal.

For each solvent, aerobic treatment was selected for managing the solvent F_input. The concentration under the BN and the FAHP suggests that biological treatment is recommended when $w^{(c)}$ is 25.50 %wt/wt for methanol; 1.00 × 10⁻⁴ %wt/wt for ammonia; 0.51 and 25.50 %wt/ wt for ethylene glycol; and 1.00 × 10⁻⁴, 0.51, and 25.50 %wt/wt for n- hexane. Additionally, F_input is wastewater for all the cases. This fact may be explained because several bioenvironmental factors affect bacteria activity and the rate of biological reactions. The biochemical oxygen demand is efficiently treated in the range of 60–500 mg/L (≈0.01 to 0.05 %wt/wt) (Samer, 2015). However, aerobic treatment suggestion for $w^{(c)}$ at 25.50 %wt/wt might be because 25.50 %wt/wt is in TRI concentration range code 1, which describes $w^{(c)}$ higher than 1.00 % wt/wt. Since $w^{(c)}$ is a crucial design and operating parameter for this treatment technology, employing values range for describing $w^{(c)}$ may be a disadvantage for applying these PAUs in these EoL flows. Another reasonable explanation for the above suggestion is the reporting facilities might have used aerobic wastewater treatment above 500 mg/L (≈0.05%wt/wt) biochemical oxygen demand by applying enough dilution rate (Samer, 2015).

As in case study one, after selecting the PAU technologies for each F_input, it is possible to allocate chemical flows downstream of each PAU. $F_{\mathit{\text{fugitive}}}^{(\text{c})}$ may be larger for F_input having ammonia when $I_{IB}^{(c)}$ = Yes, e.g., $F_{\mathit{\text{fugitive}}}^{(\text{c})}$ is 1.06 × 10⁰ kg/yr and 6.39 × 10⁻¹ kg/yr for stream # 13 and 17 (see Table S6). However, considering CV, the intervals for $F_{\mathit{\text{fugitive}}}^{(c)}$ is [0, 2.49 × 10⁰ kg/yr] and [0, 1.43 × 10⁰ kg/yr] for stream # 13 and 17. From a statistical viewpoint, both values may be the same. This result is coherent because $F_{\mathit{\text{fugitive}}}^{(c)}$ is directly proportional to the mass under analysis.

4.3. Case study 3: EoL flow from pharmaceutical preparation manufacturing

As indicated in Section 3, the first two case studies only use the information for individual PAU technologies instead of the PAU sequences found in the PAU dataset. This third case study is employed to demonstrate the framework effectiveness in designing a network of PAUs and estimating the potential releases of the chemical of concern. An existing record from the PAU dataset is randomly selected by considering only records with PAU sequences for comparing this sequence with the one designed and built by BNs and FAHP. The randomly selected record has a gaseous F_input having methanol, dichloromethane, toluene, and N,N-dimethylformamide. This record uses a PAU sequence consisting of a condenser, a scrubber, and an absorber. The record belongs to a facility in pharmaceutical IS. The problem specifications for BNs and the value for FAHP criteria can be found in the GitHub repository (Hernandez-Betancur et al., 2020b). The problem specifications used for obtaining the results are the same reported by the facility. The PAU sequence obtained following the procedure in Fig. 5 is compared with the one in the selected record; therefore, testing the framework. Fig. 6 presents the PAU sequence obtained and the CFA for each chemical.

Fig. 6.

PAU selection and sequence and CFA for the record randomly selected from the PAU dataset. F_input for this case study is 3.18 × 10⁹ kg/yr ¹ Stream # in the Flow Diagram. ² For the TRI Program, A represents air emissions. Additionally, liquid wastes having more than 50.00 %wt/wt of water are considered wastewater (W); otherwise, they are taken as a nonaqueous material (L). NA is not applicable, and it is used for destroyed or converted mass streams. ³ The bright red color represents the flows have high variance or variability, i.e., CV > 1. ⁴ All chemicals in the randomly selected record have a concentration in the interval 0.01–1 %wt/wt. For the CFA, the concentration was set at the middle of the interval, i.e., 0.051 % wt/wt.

As depicted in Fig. 6, the PAU network designed consists of a scrubber followed by a fume/vapor incinerator. Although this sequence is different from the selected record one (condenser, scrubber, and absorber), the sequence consists of treatment operations and neither recycling nor energy recovery technologies were selected. Additionally, as presented in the stream # 2 in Fig. 6, a scrubber was selected for handling methanol, dichloromethane, and N,N-dimethylformamide. It means it is possible that for the record sequence, the reporting facility used a scrubber to remove these three chemicals and the other operations for removing different chemicals from the same F_input (e.g., toluene). This inference is coherent because the TRI Program requires reporting the treatment unit or sequence regardless of whether all the units handle the reporting chemical. The BN and FAHP also suggest fume/vapor incineration as the most probable and feasible PAU technology for toluene. Fume/vapor incinerator also called thermal oxidizers are widely used for the treatment of volatile organic compounds and hazardous air pollutants like toluene. In addition, this operation unit performs best at pollutant loading of around 1500–3000 ppm (≈0.15 to 0.3 %wt/wt) (U.S. Environmental Protection Agency, 2002). Therefore, the PAU statistics for toluene suggests that for the reporting years 1987–2004, facilities frequently reduced EoL flow containing toluene by incineration. The CFA also indicates the stack emission or $F_{\mathit{\text{effluent}}}^{(c)}$ consists of 7.09 × 10⁴ kg/yr of methanol (CV = 0.13), 8.02 × 10⁴ kg/yr of N, N-dimethylformamide (CV = 0.01), 5.35 × 10⁵ kg/yr of dichloromethane (CV = 0.35), and 5.97 × 10⁴ kg/yr of toluene (CV = 0.14). As shown the CVs for the chemical flows from stack, these result values do not present a high variability (CV < 1). Unlike, the above flows, the $F_{\mathit{\text{fugitive}}}^{(c)}\text{s}$ for this case study also have a wide range of possible values. For instance, $F_{\mathit{\text{fugitive}}}^{(c)}$ for methanol from fume/vapor incinerator has a mean value of 5.23 × 10³, but its CV is equal to 1.48, which means that the result variability is high (the highest CV for this case study). Thus, the range of potential values for the $F_{\mathit{\text{fugitive}}}^{(c)}$ for methanol from fume/vapor incinerator is [0, 1.30 × 10⁴ kg/yr].

5. Conclusions

This work proposed a framework that tracks chemical flows, estimates releases, and identifies potential exposure pathways at on-site EoL industrial activities. The framework integrates multiple publicly- available databases and employs data-driven models and MCDM to develop PAU technology systems and estimates their ${\eta }^{(c)}$, CAPEX, and OPEX. Nonetheless, further data collection should be required to reduce cost data uncertainty and increase the reliability of the data-driven models or incorporating heuristic equations to estimate costs. The framework leverages the data and the literature to fill information gaps to complete the chemical flow tracking and allocate chemicals of concern inside EoL stage facilities and downstream of the PAU technologies.

Case studies based on food and pharmaceutical ISs were used to illustrate the framework application for CFA and allocation and the benefits of integrating and extending the framework with data-driven models and MCDM. The framework can support PAU technologies selection for managing chemicals of concern in EoL flows, considering the PAU effect on chemicals, and suggesting the EoL management activity sequences. However, some data mining enhancements may help find cross-year data relationships and identify sustainable PAU technologies for chemicals of concern from the PAU dataset. This effort might overcome potential data availability limitations due to reporting requirement changes from the TRI Program post-2004 (no need for reporting $w^{(c)}$). This reporting specification change might affect identifying recent industrial PAU developments and improvements. It is crucial to thoroughly analyze and enhance the framework for designing and recommending biological treatment operations to ensure accurate PAU technology predictions and inferences.

The framework enables the analysis of estimated chemical flow variabilities. In the future, the framework can help support a complete CFA at the EoL life stage for risk assessment, avoiding case-by-case studies, by connecting the on-site EoL framework with the chemical off-site tracking framework and developing and exploring robust data- driven models. The further investigation of robust data-driven models would support CFA and EoL exposure pathway characterization to extend the framework to chemicals outside the PAU dataset. Nonetheless, the models must incorporate techniques to thoroughly assess their uncertainty and reliability.

Acronyms

ASM

Annual Survey of Manufactures

BN

Bayesian Network

CAPEX

Capital Expenditures

CDR

Chemical Data Reporting

FRS

Facility Registry Service

CFA

Chemical Flow Analysis

CV

Coefficient of Variation

EoL

End-of-Life

FAHP

Fuzzy Analytic Hierarchy Process

IS

Industry Sector

MCDM

Multi-Criteria Decision-Making

NAICS

North American Industry Classification System

OPEX

Operating Expenses

PACE

Pollution Abatement Costs and Expenditures

PAU

Pollution Abatement Unit

SUSB

Statistics of U.S. Businesses

TRI

Toxics Release Inventory

VoS

Value of Shipments