Tracking end-of-life stage of chemicals: A scalable data-centric and chemical-centric approach

Abstract

Chemical flow analysis (CFA) can be used for collecting life-cycle inventory (LCI), estimating environmental releases, and identifying potential exposure scenarios for chemicals of concern at the end-of-life (EoL) stage. Nonetheless, the demand for comprehensive data and the epistemic uncertainties about the pathway taken by the chemical flows make CFA, LCI, and exposure assessment time-consuming and challenging tasks. Due to the continuous growth of computer power and the appearance of more robust algorithms, data-driven modelling represents an attractive tool for streamlining these tasks. However, a data ingestion pipeline is required for the deployment of serving data-driven models in the real world. Hence, this work moves forward by contributing a chemical-centric and data-centric approach to extract, transform, and load comprehensive data for CFA at the EoL, integrating cross-year and country data and its provenance as part of the data lifecycle. The framework is scalable and adaptable to production-level machine learning operations. The framework can supply data at an annual rate, making it possible to deal with changes in the statistical distributions of model predictors like transferred amount and target variables (e.g., EoL activity identification) to avoid potential data-driven model performance decay over time. For instance, it can detect that recycling transfers of 643 chemicals over the reporting years (1988 to 2020) are 29.87%, 17.79%, and 20.56% for Canada, Australia, and the U.S. Finally, the developed approach enables research advancements on data-driven modelling to easily connect with other data sources for economic information on industry sectors, the economic value of chemicals, and the environmental regulatory implications that may affect the occurrence of an EoL transfer class or activity like recycling of a chemical over years and countries. Finally, stakeholders gain more context about environmental regulation stringency and economic affairs that could affect environmental decision-making and EoL chemical exposure predictions.

1. Introduction

Commercial, consumer, and industrial activities, including energy recovery, recycling, disposal, and other end-of-life (EoL) activities, use chemicals that could be hazardous to human health and the environment (European Chemicals Agency, 2019; U.S. Environmental Protection Agency, 2017). As a result, risk assessment for chemicals throughout their life cycles is an important step in determining unreasonable risks, selecting a candidate based on its safety profiles to be used, e.g., as a chemical reactant, and imposing regulatory mechanisms to manage its unreasonable risks (American Chemical Society, 2019; Whittaker, 2015). Risk assessment demands collecting comprehensive information, e.g., to estimate release quantities and identify potential exposure scenarios (Organization for Economic Co-operation and Development, 2012). The data collection makes the evaluation a time-consuming task and even worse at the EoL stage because of the epistemic uncertainty about the exact pathway taken by the chemical flows (Ragas, 2011). In addition, the ever-increasing number of chemicals on the market can pose a challenge either for regulatory decision-making by government agencies or for business stakeholders assessing whether a product satisfies compliance before entering the market (European Environment Agency, 2009).

In order to assess the risk, chemical exposure practitioners must collect flow inventory data. These data, which are part of the Life Cycle Inventory (LCI), overlap the flow of materials required for life cycle assessment (Meyer et al., 2020). Researchers have used material flow analysis to get chemical flow inventories for a wide range of chemicals (Bornhöft et al., 2013; Gottschalk et al., 2010; Meyer et al., 2020; van Gils et al., 2020). Hernandez-Betancur et al. (2021b) previously developed a data engineering framework that uses publicly-accessible and siloed database systems for EoL chemical flow analysis (CFA). Environmental modelling practitioners can use this framework to rapidly perform CFA, tracking EoL chemical flows generated at industrial facilities and transferred off-site for further EoL management. The framework identifies potential EoL exposure scenarios and leverages facility-level information to estimate the flows of the transferred chemicals that the off-site EoL facility may release to the environment. Similarly, another approach goes into the EoL management facilities, identifying the pollution abatement unit technologies (e.g., solvent/organic recovery via thin-film evaporation) and performing bottom-up CFA (Hernandez-Betancur et al., 2021a). Using technology-level information improves the CFA by allocating the chemicals downstream of the abatement technologies. Hernandez-Betancur et al. (2022) combined the two previous approaches for describing the EoL management chain and extended the framework for performing CFA for the chemical circular life cycle (i.e., recycling loop). In addition, via data engineering, Hernandez-Betancur et al. (2022) developed a Markov Random Field model for describing the relationship between the entities involved in the chemical EoL management chain and recycling loop. Thus, this approach could get information about entities like the generator industry sector and EoL activity and the strength of their relationship between via the model factors or “probabilities”. These three frameworks can be leveraged to perform CFA and provide granular analysis of the potential values that can be taken by the EoL management chain and recycling loop entities. However, they only use information from U.S. databases and the latest available reporting year.

The above data engineering pipelines (i.e., sequence of connected stages) could generate a machine-readable structure to build data-driven models to associate an input variable x to an output variable y (i.e., they learn the mapping x→y), thereby predicting the behavior of the EoL management chain. However, enhancements are crucial to deploy more robust ready-to-use models in the real world. The performance of deployed data-driven models can decay over time due to data and concept drift (Raj, 2021). On the one hand, data drift occurs when the statistical distribution of x changes over time (Ackerman et al., 2020). For example, when a model is trained for EoL chemical flows of 10 – 10⁵kg/yr, but there is an increase in the annual production volume of chemicals that turns the EoL chemical flows into 10³ – 10⁶kg /yr. On the other hand, concept drift when an event changes the way x and y relate to each other (i.e., a change in the ground truth) (Lu et al., 2018). For instance, due to the green chemistry and engineering boom around 2005, companies started recycling waste solvent as an alternative to destruction via incineration (Slater et al., 2012). Therefore, to be able of handling dynamic holistic changes like environmental regulation amendments, pollution prevention, and source reduction incentives, underlying and disrupting the EoL management chain and recycling loop, it is necessary to design a data engineering pipeline to enhance and extend the above three frameworks by incorporating data applicable to other chemicals, geographic locations, and reporting years. In addition, extending the applicability domain is crucial because the information supplied by each database system strongly depends on the minimum reporting requirements, thresholds, chemicals of concern, and environmental programs the database system supports.

This work moves forward by proposing a data engineering pipeline framework to collect (Baumer, 2017), analyze, and connect regulatory databases from different countries for tracking EoL chemical flows generated by industrial facilities (manufacturing the chemical, processing as a reactant or intermediate, incorporated into formulation, mixture or reaction product, incorporated into article, and using the chemical industrially) and transferred to an off-site location for further EoL management. This advancement enables greater applicability for real-world applications to the data engineering framework developed by Hernandez-Betancur et al. (2021b), for considering the change of the regulatory restriction across different geographical locations and reporting years. The approach provides information about chemicals, generator industry sectors, and EoL activities involved in the EoL transfers, as well as reporting years, annual transfer amounts, countries, and reliability scores for reporting transfer flows. These data entries allow to describe potential EoL exposure scenarios due to off-site transfers and connect the enhanced framework with methodologies for quality assessment for LCI data (Edelen and Ingwersen, 2016). The framework adheres to the philosophy of using publicly accessible databases to be replicable and distributable to stakeholders. In addition, the framework stores the data in a machine-readable structure named PRTR_transfers database that can be used by free and open-source database management systems (Purohit, 2018). This characteristic enables the framework’s scalability for operating in the real world to be incorporated as part of machine learning operations for deploying ready-to-use models and making queries for Exploratory Data Analysis (EDA) (Raj, 2021).

The information collected by the framework could not only be used for EoL chemical risk and exposure evaluation, but also for supporting rapid LCI modelling (Cashman et al., 2016) and the use of EoL sustainability indicators (Hernandez-Betancur and Ruiz-Mercado, 2019). The PRTR_transfers data could be easily connected to other data sources in order to gather additional information that can help with understanding the potential pathways that a chemical might take across the EoL management chain as a result of factors like the national environmental stringency and the organization at a molecular level, see Section S1 in the supporting information (SI) (Hernandez-Betancur et al., 2023; Young et al., 2022). Moreover, the framework is in sync with the efforts and action plans established by national government agencies like the U. S. Environmental Protection Agency and intergovernmental bodies like the Organization for Economic Co-operation and Development (OECD) to provide public and open data to assess progress toward sustainable development (Organization for Economic Co-operation and Development, 2014; U.S. Environmental Protection Agency, 2014a). Performing data engineering, the framework moves forward to provide standardized LCI data that extends government actions and allows global-scale sustainability analysis at the EoL stage, considering progress over time, across countries, and with as much data granularity as possible.

2. Methodology

A variety of settings are required to build systems and procedures that collect, manage, and convert raw data from different siloed regulatory database systems into usable information for describing EoL aspects like the chemicals, generator industry sectors, and EoL activities involved in the off-site transfers of industrial EoL chemical flows. The above processed data should enable describing potential EoL exposure scenarios and provide LCI so that one can track a chemical to assess its potential environmental impact and exposure. In addition, these data enable linkage to additional data sources for economic information on industry sectors, the economic value of chemicals, and the environmental regulatory implications that may affect the occurrence of an EoL transfer class or activity like recycling for a chemical over years and countries.

2.1. Data sources and selection criteria

Fig. 1 depicts the general idea behind the data engineering pipeline. The data sources are the Pollutant Release and Transfer Register (PRTR) systems. In general, a PRTR is a publicly-accessible database or inventory of chemicals released to the air, water, and soil and transferred off-site for EoL management (Organization for Economic Co-operation and Development, 2014). In the PRTR systems, an off-site transfer is “the movement beyond the boundaries of the facility of either pollutants or waste destined for disposal or recovery and of pollutants in water destined for wastewater treatment”. Hence, an off-site transfer is the transport of chemicals to off-site locations for EoL activities. Member countries of the OECD must provide appropriate access to information about the environment for decision-making (Organization for Economic Co-operation and Development, 2022). Hence, members like Mexico and Colombia have implemented PRTRs following the OECD recommendations. Nonetheless, as the OECD member countries must only satisfy the minimum requirements (United Nations, 2007), the data granularity is different from one PRTR to another, depends on the country’s environmental programs supported by each PRTR, and limits data usability for environmental protection decision-making. For instance, the U.S. PRTR is considered the most comprehensive system whose data stands out due to its availability and granularity (U.S. Environmental Protection Agency, 2014a), enabling it to describe aspects like source reduction activities and on-site EoL industrial operations.

Fig. 1.

General data engineering pipeline flow for collecting, transforming, and loading the data for the PRTR_transfers database schema and records (flat icons created by Pixel perfect - Flaticon).

As presented in Fig. 1, the framework considers whether the PRTR satisfies three fundamental requirements. The first criterion is whether it is possible to access a version of the PRTR in English, facilitating its readability, comprehension, and harmonization. The second criterion focuses on whether the PRTR tracks and focuses on chemicals, not the complete waste stream, i.e., the PRTR is chemical-centric, for off-site transfers to both public sewerage systems and activities like recycling. For instance, systems like the European PRTR require reporting of information about the chemicals transferred to public sewerage facilities, but transfers to scenarios like recycling are waste-centric, so European facilities only submit whether the transfer is hazardous or not hazardous waste (Organization for Economic Co-operation and Development, 2014). Hence, the second criterion is important because the inclusion of PRTRs reporting chemical-centric data for only transfers to sewerage systems instead for all EoL transfer scenarios will lead to bias in the sample data (i.e., imbalanced data) for future data-driven models. The building of classification models is hampered by imbalanced data, potentially misclassifying EoL activities (Sun et al., 2009). The third and last criterion is the granularity of the PRTR system, enabling the identification of more specific EoL exposure scenarios. For instance, the U.S. PRTR does not only provide data for disposal scenarios but also allows one to identify whether the disposal is, for example, for underground injection or landfilling, thereby identifying and easily connecting with EoL generic exposure scenarios (Money et al., 2011).

2.2. Considerations for data harmonization

Siloed data sources describing chemicals, EoL activities, and generating industry sectors must be harmonized for consistency across the PRTR systems selected for building the PRTR_transfers database. For chemical substances harmonization, it requires to keep a unique record for identifying the chemical inside the resulting database and across different siloed database systems like PubChem. Thus, a primary key is assigned to each chemical inside the PRTR_transfers, as well as the numerical designation maintained by the Chemical Abstract Service (CAS) is kept. The CAS number facilitates access to chemical price information and cheminformatics descriptors (Heller et al., 2015; Weininger, 1988). When a CAS number cannot be retrieved through the PRTR system, the framework connects to services like those used by the U.S. Substance Registry Service to retrieve it. This is done by utilizing the chemical name registered in the PRTR program (U.S. Environmental Protection Agency, 2019). A manual search is carried out if the number cannot be located automatically. As PRTR systems provide information about chemicals belonging to groups, the framework accesses lists containing information about chemicals of concern to identify hazardous and commercial chemicals that can belong to groups like chlorophenols and polycyclic aromatic hydrocarbons. Lists like the one provided by the Environment and Climate Change Canada (2017a), National Research Council Committee on Pyrene and Selected Analogues (1983), and the U.S. Environmental Protection Agency (2005) are used.

Each PRTR system has a unique way of classifying EoL activities or EoL transfer classes, therefore, to create a PRTR_transfers harmonization system for the EoL classes, these categorizations are extensively investigated and compared. In this paper, the PRTR systems that have reporting records for each EoL transfer class are chosen. As a result, while greater granularity may be sacrificed, bias and database imbalance are addressed, avoiding the introduction of counterproductive effects into serving data-driven models. An example of the harmonization procedure for the EoL transfer classes is the U.S. PRTR system differentiates between the wells suitable for injecting hazardous wastes as underground injection class I wells, while the Canadian PRTR only has underground injection category for both hazardous and non-hazardous wastes. Thus, the PRTR_transfers database provides underground injection no matter whether the chemical is considered hazardous. In addition, transfers to public sewerage systems are kept as a separate category, no matter if they are for treatment or disposal. Finally, the industry sector classes in the PRTR_transfers are proposed based on the 4th revision of the International Standard Industrial Classification (ISIC) that enables cross-walking of the industry classification systems established by each country (United Nations, 2008). The use of ISIC enables data enrichment by easily connecting with environmental and economic data of the industry sectors to beneficially gain context for decision-making, analysis, and prediction, considering temporal and geographical correlation. ISIC follows a 4-level hierarchical structure that includes sections, divisions, groups, and classes, with the sections at the top level (less specific) and the classes at the bottom one (more specific). For the PRTR_transfers, the framework utilizes the ISIC divisions (around 90) for clustering the industry sectors. This selection considers a trade-off between data specificity and sample availability to avoid not including samples containing information for certain industry sectors, which could lead to decision-making bias.

2.3. Considerations for data extraction, storage, accessibility, and reliability

Just like other data sources, the PRTRs could be either structured (e. g., databases), semi-structured (e.g., web pages), or unstructured (e.g., documents). Hence, the framework pipeline extracts information from those sources no matter their structure. Following extraction, the data is transformed into a generic and tabular form, structuring the PRTR_transfers in accordance with normal forms of relational database theory, reducing data redundancy and improving data integrity (Kent, 1983). Finally, the transformed data are stored in a free and open-source database management systems, guaranteeing PRTR_transfers scalability and accessibility (Krocz et al., 2020). In addition, each record in the PRTR_transfers assign a reliability score based on guidelines for quality assessment for LCI data proposed by Edelen and Ingwersen (2016), which is like the pedigree matrix applied by ecoinvent (Ciroth et al., 2016) and widely used by LCI practitioners (Vélez-Henao et al., 2020). Hence, every basis of estimate for the chemical flow transfers reported by facilities to the PRTR programs has been mapped to a flow reliability score according to its verification or validation method. The number 1 is assigned to verified data based on measurements; the number 2 represents verified data based on a calculation or non-verified data based on measurements; the number 3 represents non-verified data based on a calculation; the number 4 represents a documented estimate; and the number 5 represents an undocumented or estimate.

3. Results and discussion

After a thorough review of the existing PRTR systems belonging to the OECD member countries, only three out of 7 (including the PRTR for European Union Member States or E-PRTR) were selected based on the criteria described in Section 2.1. The selected siloed systems are the National Pollutant Release Inventory (NPRI) of Canada (Environment and Climate Change Canada, 2017b), the Toxics Release Inventory (TRI) of the U.S. (U.S. Environmental Protection Agency, 2013), and the National Pollutant Inventory (NPI) of Australia (Department of Agriculture, Water and the Environment, 2001). This section is composed of three subsections to discuss the impact and limitations of this work for future developments. Section 3.1 introduces and provides the PRTR_transfers database structure and how to use it for connecting with LCI endeavors. Section 3.2 shows the EDA performed to obtain insights and implications for the development of future data-driven models, e.g., for predicting how probable is an EoL transfer scenario for a chemical of concern. Finally, Section 3.3 gives a general comparison of the current framework and the one proposed by Hernandez-Betancur et al. (2021b). The Python scripts written to execute the data engineering pipeline and obtaining the PRTR_transfers database can be found in a public GitHub repository, see Section S3 in the SI. In addition, readers can explore more about the industry sectors, chemicals, and transfer classes in Section S5.

3.1. PRTR_transfers structure

The PRTR_transfers database contains a total of 3116,211 records and its reporting years ranging from 1987 to 2020. The records in PRTR transfers, which correspond to 72 out of the 90 ISIC divisions (e.g., extraction of crude petroleum and natural gas), were obtained using the proposed data engineering pipeline. A total of 643 harmonized compounds, e.g., m-cresol and chlorine oxide, have data in the PRTR_transfers database. Additionally, there are ten EoL activities or transfers classes that, according to the waste management hierarchy (i. e., recycling, energy recovery, treatment, and disposal), can be further aggregated into five groups. surface impoundment, landfill, storage, underground injection, and other disposals are included in the disposal group. Destruction and other treatments are included in treatment. Only one transfer class is present in the recycling and energy recovery groups. The final group relates to sewerage with a single transfer class. Sewerage refers to such EoL transfers to publicly owned wastewater treatment facilities. It is taken into consideration individually despite not being officially addressed in the waste management hierarchy due to regulatory applications and evaluations. (U.S. Environmental Protection Agency, 2017).

Fig. 2 presents sample data obtained from the PRTR_transfers database containing information for ethylbenzene transferred to off-site locations by facilities in the industry sector associated with manufacture of chemicals and chemical products in Canada during 2019. Fig. 2 solely shows data for recycling, landfill, and energy recovery for the sake of simplicity. The framework data make it possible to connect the sectors of the generating industries to probable EoL activities for the movement of ethylbenzene material to off-site places. For instance, the annual transfer of ethylbenzene for recycling ranges from 288 to 18,223 kg, with an average of 4608 kg and a median of 1770 kg. Like ethylbenzene, stakeholders can utilize the PRTR_transfers database to develop data-driven models as outlined in Section 3.2 or to collect data to complement their LCI at the EoL stage.

Fig. 2.

Sample data for off-site ethylbenzene transfers reported to the Canadian PRTR system (i.e., NPRI) by facilities belonging to the manufacture of chemicals and chemical products industry sector during the reporting year 2019.

Additionally, the PRTR transfers database data can be linked to the pedigree matrix used by ecoinvent because it contains the reliability score considering guidelines for LCI data quality (Ciroth et al., 2016). As presented in Fig. 2, PRTR_transfers database only supplies the data for off-site EoL transfers of chemicals. Nevertheless, industrial facilities’ chemical discharges to the soil, water, and air might be collected using the harmonization system described in the manuscript. As discussed in Section 3.3, the information gathered by this suggested data engineering can be utilized to describe the pollution abatement techniques applied to handle chemicals of concern, the EoL supply and management chain, and data on what occurs inside industrial facilities. See Section S2 in the SI for more details on the PRTR_transfers schema and structure.

3.2. Exploratory data analysis and implications

Reproducible LCI data require data quality assessment. As mentioned before, this work incorporates the methodology developed by Edelen and Ingwersen (2016) for assigning a reliability score to the reporting chemical transfer flows. Fig. 3 presents the distribution of the reliability score of the PRTR_transfers records. As illustrated in Fig. 3, 40.32% of records have mostly a score of 4, i.e., they are data records whose data generation method is documented estimate. In contrast, 50.39% is data whose score ranges from verified data based on measurements (1) to non-verified data based on a calculation (3). The data’s reliability score varies for each of the three nations in the PRTR_transfers database, as would be expected. Some observations include the fact that no record from Australia has a 5 score (either recorded or estimated), 42.92% of Canadian records have a 5, and 42.26% of US recordings have a 4 (see Fig. S2 in the SI). In the data-driven modelling stage, one can use only those reliable records based on their scores and even drop duplicate records with lower reliability, thereby ensuring a better and more reliable model performance. Removing or not removing a record with a specific score depends on the modelling criteria established by the stakeholders, the desired accuracy according to the project scope, and the chosen level of automation for the predictive production system (Wang et al., 2021).

Fig. 3.

Distribution of the reliability score for the 3116,211 records loaded by the data engineering pipeline to the PRTR_transfers database.

Class imbalance is a factor that may impair the performance of upcoming data-driven models; as a result, EDA must offer information on this potential problem that may affect how environmental modelling practitioners use PRTR_transfers data. When it comes to PRTR_transfers, the class imbalance problem can arise if some transfer classes have a significant excess of records (known as majority class) relative to others (known as minority class), e.g., more records for disposal operations than for recycling (Luque et al., 2019). Fig. 4 shows how the records are distributed in each EoL transfer class. Purple is for classes that are overrepresented by the data, and blue for the underrepresented ones. Fig. 4a outlines the situation if recycling, sewerage, landfill, energy recovery, other treatment, other disposal, destruction, storage, underground injection, and surface impoundment are used as the target variable values for a classification algorithm. For the minority EoL transfer classes like surface impoundment, the model may misclassify, pointing out the EoL scenario would not be this class when it does is (type II error or false negative). The contrary may happen for the majority classes like recycling, where the algorithm misclassifies, saying that recycling would occur when it does not (type I error or false positive) (Banerjee et al., 2009). Such circumstances would be problematic because assuming that an EoL scenario for a hazardous chemical is safe would prevent an industrial facility from establishing an effective risk control, which might result in an overexposure that would have detrimental effects on the human health and the environment (U.S. Environmental Protection Agency, 2014b), or spending unnecessary time and resources further analyzing an EoL scenario or supposedly mitigating a risk that never existed.

Fig. 4.

Distribution of EoL transfer classes in the PRTR_transfers database for the 3116,211 records. (a) imbalance analysis by each of the 10 EoL transfer classes in the database. (b) imbalance analysis by each of the EoL activities (recycling, energy recovery, treatment, disposal, and sewerage). If the class is a minority, it is blue; if it is a majority, it is purple; and if otherwise, it is green.

A different context occurs if recycling, energy recovery, treatment, disposal, and sewerage were used as the target variable values (see Fig. 4b). In this case, disposal would be the majority class, while energy recovery would be in the minority class. Nonetheless, the use of these 5 classes instead of the other 10 would lead to a lower imbalance degree for the sample data. No matter what classes one selects as the target variable values, users should either drop those minority classes, implement any data under-sampling, or over-sampling technique to resample and balance the training dataset before modelling (Spelmen and Porkodi, 2018). In addition, Fig. 5 presents the distribution of the 10 EoL transfer classes by country. Fig. 5 schematizes how different the distributions of the transfer classes are for the three countries in the PRTR_transfers. For instance, facilities in Canada have reported a greater implementation of recycling activities (29.87% of its samples) than in Australia (17.79% of its samples). According to the OECD statistics, Canadian environmental policies have been more stringent than in Australia and U.S. since 2003 (Organization for Economic Co-operation and Development, 2016). Hence, the above can point out there may be a correlation between the environmental stringency and the distributions depicted in Fig. 5, which could be leveraged by considering across-country information.

Fig. 5.

Distribution of the 10 EoL transfer classes in the PRTR_transfers database for 3116,211 records by country.

As mentioned before, data drift may occur, so the statistical distribution for x variables (e.g., the amount of chemical transferred) changes over time, thus affecting the performance of data-driven models and their predictions. To show this concept graphically, transfer flows for recycling and destruction in the reporting years 1987, 2005, and 2019 are filtered. Those years were chosen because, around 2005, a particularly strong influence of green chemistry and engineering philosophy began, with a particular emphasis on replacing destruction activities by recycling alternatives (Slater et al., 2012). A violin plot is used to show changes in the probability density of the data at different values. In the violin plots in Fig. 6, readers can look at the distribution of destruction and recycling flows on a logarithmic scale in 1987, 2005, and 2019. The first thing to observe is the distribution of each transfer class is different from one year to another, proving the existence of data drift across the reporting years. The second item is the highest value of distribution for destruction was greater in 1987 than in 2005 and 2019. The contrary occurs for the distribution of recycling, where the peak is lower in 1987 than in 2005 and 2019. Hence, the proposed data engineering pipeline would strengthen the machine learning operations in production due to supplying labeled data to the ready-to-use model if retraining is required (Raj, 2021). The data supply rate can be one year because of the PRTR system reporting cycles, which is reasonable considering the underlying phenomena that can affect and disrupt the EoL management chain do not vary steeply (e.g., environmental regulation amendments and manufacturing volume).

Fig. 6.

Violin plots representing the distribution of the destruction and recycling transfer flows on a logarithmic scale. These aggregate the data for the three reporting countries. The plots contain data for 1987, 2005 (the year around which the green chemistry and engineering boom happened), and 2019.

The other aspect that could occur with a data-driven model in production is concept drift, i.e., a change in the ground truth, which leads to the performance of the ready-to-use model decaying over time. A shift in the ground truth indicates that industrial facilities have altered their willingness to implement an EoL activity over time due to drivers like advancements in technology, environmental regulations, and environmental awareness. To depict and analyze graphically this concept, the bar plot in Fig. 7 presents the distribution of the 10 EoL transfer classes in 1987, 2005, and 2019. As shown by the difference in the bar heights, a change in the distribution of transfer classes has occurred across the reporting year. If the transfer class is chosen as the target variable for a classification algorithm, the deployed model might misclassify due to a change in ground truth or concept drift. For instance, the highest bar in the picture corresponds to sewerage in 1987. So, if someone had built that model in 1987 without a continuous data supply, the ready-to-use model would have misclassified the 2019 ground truth, probably incurring a type I error or false positive, where the most probable scenario could have been recycling (the highest bar in 2019). Therefore, the framework will supply data at an annual rate, for tracking changes in the relationship for x→y and for retraining data-driven models if required, keeping a data-centric paradigm, i.e., iteratively improving the data, and holding the model fixed (Ng, 2021).

Fig. 7.

Distribution of transfer classes in 1987, 2005 (the year around which the green chemistry and engineering boom happened), and 2019. The bar plots aggregate the data for the three reporting countries.

The PRTR_transfers database can be used to identify the top ten industry sectors that reported the highest total amount of chemical substances transferred during the reporting years. Thus, industry sectors associated with activities like “mining support service activities” (sector 9 in Fig. 8) and “waste collection, treatment and disposal activities; materials recovery” (sector 10 in Fig. 8). The stacked area plot in Fig. 8a shows the sector 10 has transferred chemical pollutants in quantities ranging from a 500 million kg/yr. to around 2.5 billion kg/yr. (with a peak in 1999). In addition, as presented in the heatmap in Fig. 8b, the mass flow has been mainly transferred by this sector to recycling, followed by energy recovery. Fig. 8b shows the mass flow shipped to recycling is substantial for the other industries in the top ten. Thus, if Figs. 4a and 8b are analyzed jointly, they indicate that recycling scenarios do not only represent 21.26% of the samples but also a representative amount of the total cross-year transfers. In addition, Fig. 6 points out the central tendency measurement of the distribution (i.e., the mode) associated with the individual recycling transfers is above 1800 kg/yr.; moreover, the three distributions present negative skewness, which means the probability distribution is especially concentrated on high recycling flow transfers.

Fig. 8.

Total amount transferred for the top ten industry sectors. These plots aggregate the data for the three reporting countries. (a) Total transfer amount by industry sector from 1987 to 2020. (b) distribution of the total transfer amount among the transfer classes based on industry sector.

The appearance of a chemical substance in a PRTR system depends on regulatory needs. Thus, a substance can be in a PRTR but not in others, or even be reported in a program from, until, or between certain years. Fig. 9 presents six chemicals (out of 643) selected to demonstrate the usefulness of cross-year and cross-country data to increase the applicability domain of data-driven models, especially regarding an environmental stringency context and molecular descriptors. For example, Fig. 9f shows that dichlorofluoromethane was only reported to the U.S. PRTR (i.e., TRI) from the reporting period between 1995 and 2002. Carbon monoxide (in Fig. 9d) and methyl carbitol (in Fig. 9e) have only been reported to the Australian and Canadian PRTRs, respectively. There are substances like 2-butanone (in Fig. 9c) that stopped being reported to the U.S.’s PRTR but started being considered by the other regulatory reporting systems. In addition, there are substances whose reporting flow transfers could change their flow distribution from one country to another, like hydrochloric acid (in Fig. 9a) and carbonic dichloride (in Fig. 9b), which have reported higher flows for Australia and Canada than in the U.S., respectively. In short, the inclusion of other chemical-centric PRTR systems may represent a benefit for the domain where a data-driven predictor ranges. However, an outlier detection technique should be used to ensure the data-driven model can learn and not harming its performance (Zharikova et al., 2022).

Fig. 9.

Total transfer amount across reporting years by chemical substance. The X-axis represents the reporting year, and the Y-axis represents the total transfer amount in kg/yr. (a) hydrochloric acid, (b) carbonic dichloride, (c) 2-butanone, (d) carbon monoxide, (e) methyl carbitol, and (f) dichlorofluoromethane.

Environmental modelling practitioners can create predictive models using the PRTR_transfers data that, for instance, offer the likelihood that a chemical may have EoL activity after being transferred (classification problem) or forecast the amount of chemical transferred for an EoL activity (regression problem). Thus, how someone decides to use the data depends on the purpose of the project established at the beginning. Furthermore, another advantage of using cross-year data is the PRTR_transfers information could be connected to information about the environmental stringency over the years, which could represent the current reality of other countries. Similarly, PRTR_transfers could be linked with economic data to know the industry sector value change over the reporting years, mimicking the reality of the same sector in different countries like Brazil. The above two facts make possible data enrichment for gaining insights for decision-making about the chemical EoL management chain and increase the applicability domain of data-driven models. Readers can visit the website listed in Section S4 of the SI to look at additional relationships that may have an impact on whether a chemical will exhibit any EoL activity. This website has dynamic dashboards that allow for this analysis.

3.3. Comparison and connection with the previous methodologies: scalability

Hernandez-Betancur et al. (2021b) created a framework for tracking off-site EoL transfers with a data granularity greater than the PRTR_transfers database for the transfer classes and industry sectors. That framework uses around 32 EoL transfer classes belonging only to the TRI, as well as the industry sector classification from the U.S. Census Bureau (U.S. Census Bureau, 2017). Nonetheless, its granularity might lead to underrepresenting some transfer classes and industry sectors, causing false positives and false negatives, as shown in Section 3.2. In addition, transfer classes like RCRA Subtitle C landfills are only of interest under the U.S. environmental regulation of hazardous wastes (specifically the U.S. Resource Conservation and Recovery Act) (U.S. Environmental Protection Agency, 2010). Although the industry classification system used by Hernandez-Betancur et al. (2021b) agrees in structure and composition with the one used by Canada and Mexico, there are exceptions to the rule to meet national needs (Statistics Canada, 2002). In contrast, the PRTR_transfers database offers a transfer classification that, although it is less granular, represents the general interest of different countries, especially for rapid EoL exposure scenarios screening. The developed approach is rapidly cross-walking to other industry classifications used by different countries members of the OECD, like Germany and France, by using ISIC and its industry classification as a reference. Thus, the enhanced framework takes the underlying idea proposed by Hernandez-Betancur et al. (2021b) and extends it to incorporate useful data beyond the U.S. database systems, making it easily applicable to other geographical locations and years so that it can be used for LCI data, CFA, and building data-driven models for the real-world (see Fig. 10).

Fig. 10.

The relationship between the suggested framework and earlier approaches, as well as its potential application to data-driven modelling.

As previously mentioned, the framework can link siloed chemical-centric PRTR systems annually to track EoL chemical flows that are shipped to off-site destinations. Those flows are produced by industrial facilities when they manufacture the chemical, process it as a reactant or intermediary, combine it into a formulation, mixture, or reaction product, incorporate it into a product, and use the chemical industrially. As shown in Fig. 10, the database created by the data engineering pipeline can be used to create data-driven models and train them annually, avoiding the issues described in Section 3.2 related to changes in the statistical distribution of the variables that could have an impact on how businesses use an EoL activity. The PRTR transfers database design allows environmental modelling practitioners to connect with other data sources, such as OECD statistics to gain insights on the environmental stringency in the OECD member countries and the annual gross value added by industries in the member countries, as well as other sources containing parameters like chemical prices (associated with the value of a chemical) that can influence the occurrence of an EoL activity. This is illustrated in Fig. 10 as another crucial aspect.

Fig. 10 also shows how the data in the PRTR_transfers database might be linked to the framework created by Hernandez-Betancur et al. (2021a). An environmental modelling expert can use this relationship to determine the pollution abatement unit technologies that are connected to the EoL transfer classes in the PRTR transfers database, as well as potential industry sectors that receive and manage EoL activities for chemicals of concern. So, CFA can be performed at the waste management plant itself. The PRTR_transfers data can also assist in the development of data-driven models that produce the factors or “probabilities” that connect the generator industry sector and EoL activity entities in the Markov Random Field proposed by Hernandez-Betancur et al. (2022) to describe the EoL management chain and recycling loop. The development of a multi-level and multi-scale modelling system based on Hernandez-Betancur et al. (2021a), Hernandez-Betancur et al. (2022), and this work, however, should be the main focus of future research.

4. Conclusions

This work contributes to developing a data-centric and chemical-centric framework for tracking EoL chemical flow transfers, identifying potential EoL exposure scenarios, and performing CFA. The data engineering pipeline leverages publicly-accessible PRTR systems belonging to OECD member countries. The LCI data obtained by the approach is stored in a database called PRTR_transfers that could be connected to MLOps in production environments, making the framework scalable for real-world applications. The data ingestion pipeline can supply data at an annual rate, ensuring labeled data can be ingested into data-driven models if retraining is needed, especially to face problems like data and concept drift that could drastically affect the performance of data-driven models.

Although the framework has less resolution about the industry sectors and EoL transfer classes compared to previous developed methodologies, it can support the rapid screening of EoL potential scenarios by reducing the time consumed for collecting comprehensive data for performing such a task. In addition, cross-year data ingestion can lead to increased applicability domain for data-driven models and enrich the context by ingesting additional data about environmental policy stringency across the years and geographies as well as economic insights about, e.g., economic value added by industries through reporting years and the perceived value of a chemical substance by the market.

Furthermore, the developed framework could be connected to previous data-driven approaches for performing CFA of pollution abatement unit technologies and understanding the potential behavior of the chemical EoL management chain and recycling loop. Nonetheless, more work should be done in future developments to connect those methodologies for creating scalable multi-level and multi-scale modelling systems to rapidly screening potential chemical exposure scenarios on the premise of conserving a data-centric paradigm. The results and finding from this work can be a starting point to change the reporting requirements established by the regulatory databases, thereby incorporating the needed data entries for LCI and CFA at the EoL stage.

Acronyms

CAS

Chemical Abstract Service

CFA

Chemical Flow Analysis

EDA

Exploratory Data Analysis

EoL

End-of-Life

ISIC

International Standard Industrial Classification

LCI

Life Cycle Inventory

NPI

National Pollutant Inventory

NPRI

National Pollutant Release Inventory

OECD

Organization for Economic Cooperation and Development

PRTR

Pollutant Release and Transfer Register

TRI

Toxics Release Inventory

Data availability

Data will be made available on request.