Introduction

In its “Climate Change 2021: The Physical Science Basis” report, the Intergovernmental Panel on Climate Change (IPCC) laid out a clear and grim prediction of the future climate (1). In even the lowest greenhouse gas emissions scenario considered, it is more likely than not that a 1.5°C global warming level will be exceeded in the mid-21st century. And only in the lowest emissions scenario does that level fall back under 1.5°C by the century’s end.

Meeting this lowest emissions requirement, in line with the Paris Agreement pledge (2), requires unprecedented transformation across all facets of society. The chemicals industry has a vital role to play in enabling this transition, by reducing emissions of processes currently using fossil feedstocks, and transitioning to emerging carbon feedstocks such as waste, biomass and carbon dioxide.

Several major industrial players have made net zero pledges (3) and the magnitude of this task should not be overlooked. Based on the International Energy Agency’s recent roadmap to net zero by 2050 (4, 5), the CO₂ emissions of the chemical industry must fall by an approximate average of 9.1% year-on-year between 2020 and 2050 to reach net zero targets, a substantial step-up from the 1.4% year-on-year decrease under current announced pledges (5). These data are summarised in Figure 1.

Many chemical companies are unsure how to meet their net zero pledges, highlighted by a recent survey from Black & Veatch, USA (6). This survey showed that while 80% of companies with revenues over US$250 million set greenhouse gas reduction targets, 25% of them set goals at such a level that they were unsure how to meet them.

While this task is daunting, it is not impossible. In fact, there are many possible decarbonisation pathways proposed by the chemicals industry, such as the use of bio-based feedstocks, carbon capture and utilisation (CCUS), electrification of processes using renewable electricity, incorporation of low carbon hydrogen and more (7). The optimum solution varies for different chemical processes, different regions and different infrastructure, so selecting which technologies to invest and innovate in at an early stage is a challenge in of itself. This is also a key consideration when considering what incentives, credits or funding schemes are applicable to support development and implementation.

To address this challenge, the chemicals industry requires data-driven analysis approaches to help inform early-stage decision making. These approaches should be accessible to a diverse group of stakeholders (technologists, project developers, financiers), flexible to the wide array of different process solutions available, and quick enough to adapt to the fast-changing industry.

Life Cycle Analysis

“All models are wrong, but some are useful” is an appropriate aphorism when discussing sustainability assessments (8). The most common form of process and product environmental sustainability assessment is LCA. Considerable effort has been made to standardise LCA methodologies and these offer the most accurate estimates for carbon intensity of products. They are of course still only models, so many assumptions must still be made. The main issues with the LCA approach for this work are its site and process specificity and its labour-intensity.

While LCA databases like ecoinvent^TM (9) provide equivalent CO₂ figures for various chemical products such as plastics, this data lacks granularity. The specific process steps used are often unclear and it is non-trivial to alter specific steps. Finding sources of data points takes considerable time (with full LCAs taking weeks or months to complete) and interacting with the data through programs such as SimaPro^TM (10) can be non-intuitive.

The authors would like to be clear that the time-consuming nature of LCA processes is necessary to allow for adequate quality control and the involvement of external practitioners, both of which are critical to prevent methodology misuse and deception. However, to meet these necessary standards the full LCA process must be repeated for every path from raw material to final product, which is unfeasible for early-stage, high-level investigations. And so, a complementary and alternative tool to guide early pathway thinking and innovation decisions is highly desirable.

The Chemical Networks Approach

A novel approach, as discussed here, involves invoking the principles of graph theory to model chemical processes. Graph theory is a complex field of mathematics and will not be covered in detail here. However, a brief introduction of terms is presented in the Supplementary Information accompanying the online version of this article.

The abstract concept of graph theory underpins networks that are essential to our day-to-day lives, from the internet to satellite navigation. Networks have previously been used to describe chemical reaction phenomena, notably for the dynamic behaviour of biochemical systems (11–13).

Graph theory may also be used to explore chemical processes, where chemical feedstocks and products are nodes, linked by chemical reactions or processes. In doing so, process data (and the associated sustainability metrics) may be linked, allowing a ‘big picture’ view of complex, multi-step process routes. Chemists and chemical engineers intuitively link the feedstocks and products of these routes, so it makes sense to link the data in the same way.

A concept of chemical networks was taken from conception through to commercialisation by Grzybowski et al., who created a vast chemical network of organic chemistry from the Beilstein Database (14, 15). Investigating the topology of this network elucidated some interesting insights into its structure, but the data lacked sufficient detail for more interesting questions, such as where to find novel and more efficient retrosynthetic pathways, to be answered.

Subsequent research (16) plus considerably more data and complex algorithms allowed the team to demonstrate the concept’s real-world application with predicted and validated one-pot reactions (17), optimised synthesis pathways (18) and management of chemical threats (19). The resulting ‘chemical internet’ allowed for considerable synthetic insight, with the associated Chematica program used to autonomously design synthetic pathways to medicinally relevant targets (20, 21). This software was subsequently purchased by Merck, where it exists as the purchasable SYNTHIA^TM software package (22).

While this is an impressive body of work, SYNTHIA^TM’s prime use is for the complex synthetic organic chemistry of the pharmaceutical industry. The same conceptual framework may also be used for evaluating large-scale industrial chemical processes in the commodities space. The methodology described herein was developed independently but shares the same base premise and philosophy.

In this work, industrial chemical feedstocks and products are represented as nodes in the chemical network, with the edges between the nodes containing industrial chemical process data provided from various literature, market or commercial sources. This concept is illustrated in Figure 2. These edges may be weighted by industrially relevant data such as CO₂ equivalence (CO₂e), water usage or feedstock consumption.

Similar methodologies for evaluating sustainability metrics have been employed in the past, using linear programming and ‘superstructure’ network analysis to evaluate the chemical industry (23–29). The work herein differs in its approach, in designing an explorable network of every possible route and an interactive user interface to allow users to explore all possible routes to all products in real time, as well as crude approximations of the effects of different geographies.

The analysis carried out aims to apply the principles of graph theory to industrial chemical process data in order to rapidly assess the environmental metrics of different pathways from raw material to final chemical product. These methods are then compared with existing assessment methods. In this research, the plastics production value chain was chosen as a case study. This was chosen because emissions from the plastic industry are sizeable (potentially reaching 15% of the global carbon budget by 2050 (30)), with a large proportion of industrially relevant processes involved from raw material to final product.

Materials and Methods

The first step required to represent the sustainability metrics of chemical process data as edge weights in a network is to define the simplified parameters used to represent the metrics. The goal here is to apply universal and quick-to-compute rules to all processes. Such simplification admittedly reduces the accuracy of the data provided and therefore should only be used for preliminary screening and comparison of different process routes, rather than for detailed sustainability data for an individual process or process routes.

Defining Equivalent Carbon Dioxide

This methodology defines CO₂e of a chemical process using three parameters, namely ‘direct process’, ‘direct utilities’ and ‘indirect utilities’ which are defined in detail in the following section and shown in Equation (i). These definitions are similar to those presented by IHS Markit in its chemical footprint methodology (31). A fourth parameter, relating to the CO₂ equivalence of sourcing the feedstock, is presented for discussion and further work.

​​​

(i)

where m_{target product} > 0.

The first parameter discussed is ‘direct process’. Direct process represents any CO₂ that originates from the feedstocks. It is calculated from the sum of the net equivalent mass of CO₂ formed per mass of all carbon containing useful products, subtracted from the net equivalent mass of CO₂ formed per mass of all carbon containing feedstocks, assuming complete oxidation to CO₂. This is represented in Equation (ii):

​​​

(ii)

where m_{target product} > 0, if component = feedstock, m_feedstock > 0, if component = useful product, m_{useful product} < 0.

For a defined chemical component, with a defined molar mass, calculating the CO₂e is trivial, as shown in Equation (iii):

​​​

(iii)

For more ill-defined chemical feedstocks, which consist of mixtures of chemicals, CO₂e may be estimated by approximating as one component (for example, pyrolysis gasoline may be reasonably approximated as benzene due to its high aromatic content); a mixture of defined components (for example, BTX may have a known percentage of benzene, toluene and xylenes, from which CO₂e can be estimated); or from literature (for example, coal’s carbon content will vary, so literature values can be used to give an approximation).

Exceptions may be made for bio-based feedstocks. As the CO₂ used to form these feedstocks is directly removed from the atmosphere via photosynthesis, CO₂e_{bio-feedstock} = 0. The same approach is taken with direct air capture CO₂. Further complexity may be introduced by including the CO₂ contribution of sourcing the raw materials, which is particularly important for bio-feedstocks. This will be discussed in a later section.

All useful products, defined as products which have sufficient value as to be taken as a negative cost in the process economics, are factored in, even if they are used in the process. As an example, fuel gas is produced in several petrochemical processes and then burned to cover the process heat requirements. The m_{useful product} value here is still negative, as the direct utilities CO₂e is a gross value, so will cancel this out.

The second variable is ‘direct utilities’. Direct utilities represents any CO₂ that originates from onsite combustion of fuels to generate heat, or conversely any heat energy exported from exothermic processes. Value represented in process data must be converted to a useable CO₂e. Typically, the values are split into two variables: ‘fuel consumption’ and ‘steam consumption’. Fuel consumption is defined in units of energy per mass of target product, so the CO₂e value can be calculated by multiplying this energy by the CO₂e value for the regional energy source (Table I), which is estimated from the US Environmental Protection Agency. Steam consumption is defined in units of mass per mass of target product. The energy required to generate this steam is calculated from the latent heat of saturated steam at the given pressure, multiplied by the mass of steam consumed. The CO₂e is then calculated (Equation (iv)) from the energy according to the regional values supplied in Table I.

​​​

(iv)

where m_{target product} > 0.

The third value is ‘indirect utilities’, where CO₂e is calculated in a similar way as direct utilities (Equation (v)), except here the utilities energy (for electricity and refrigeration in the process) is supplied from offsite power. Again, this is region-specific, depending on the energy mix in the regions. This is taken into account by employing the CO₂ coefficients shown in Table II. These coefficients account for differences in regional power mixes.

​​​

(v)

where m_{target product} > 0.

As mentioned earlier, adding a fourth variable representing the CO₂ equivalent of sourcing the feedstock would improve the accuracy of the estimated total CO₂e. This variable is tentatively titled ‘indirect process’ or ‘indirect material sourcing’. For example, chlorine (which has no direct process equivalent CO₂) would have a value of ~1 tonne_CO2e tonne^–1 due to the energy requirements of the chloralkali process (38). The addition of this variable would be especially helpful when assessing bio-based processes, where the carbon footprint associated with land use for growing the biomass plays an important role. Data can be found by consulting the databases available for LCA assessment (9).

The largest degree of uncertainty in these data, assuming that the process data input is accurate, is in assigning a CO₂e value to ill-defined chemical feedstock compositions, such as coal. Literature searches reveal highly variable values for the carbon density of coal, naphtha, crude oil and other raw materials. More accurate data, tabulated in a database and perhaps regionally specific, would help in defining the direct process carbon equivalence. It is helpful to use a small number of values for the carbon intensity of these feedstocks (rather than making them process specific), as this allows for easier comparison between processes and simplifies the chemical networks. The accuracy of this data is sufficient for preliminary screening and comparison of several routes based on the same feedstock.

As an example of a reasonable compromise, in this study synthesis gas (syngas) has been simplified into ‘Syngas 1:1’, ‘Syngas 2:1’ and ‘Syngas 3:1’, representing the different ratios of hydrogen and carbon monoxide. Though different processes will produce different compositions of gas, grouping in this way allows for easier comparison in chemical networks. These approximate ratios are cited in process report data, so the assumption is unlikely to add significant uncertainty to results. It should be noted that other gas products produced, such as methane, are explicitly accounted for.

Defining Other Variables

Other variables, such as the E-factor, energy and water consumption, are defined in the Supplementary Information.

Building the Network

The data for individual process steps can be built into network graphs and the chemical pathways created are used to estimate CO₂e for a particular chemical pathway. Details on how the network is built, and the rules applied, can be found in the Supplementary Information.

In order to assess the large amount of industrial chemical processes required, IHS Markit Process Economics Program (PEP) reports were primarily used for process data (39). A table of processes cited, with their associated primary feedstock, primary product and reference, are available in the Supplementary Information.

By applying these rules to a database of process data, linking each process by its chemical feedstocks and products, every possible chemical pathway from a feedstock to product may be found, and the CO₂e for this pathway can be calculated. The workflow is summarised in Figure 3, with the process data transferred to a database, the chemical network formed from the data and the individual pathway data extracted from the network.

Expanding the network is trivial as new process data may be entered into the database and the network expanded by the algorithm of linking processes by feedstocks and products.

Results and Discussion

Visualising the Network

An illustration of the chemical network corresponding to the plastics production value chain is provided in Figure 4. In this graph, each node is a chemical feedstock or product and the edges between each node are weighted by the equivalent CO₂ of the lowest carbon route from that feedstock to 1 tonne of the next product. Equivalent CO₂ from other sources such as materials sourcing, transport and leaks are not included. The selected region is the US Gulf Coast (USGC), with direct utilities supplied via natural gas and indirect utilities sourced from the regional energy mix. No special accounts of renewable energy use are used. The graph is constructed from over 200 processes, creating over 23,000 possible routes to the final polymers. The process data are sourced from literature articles, internal analysis and IHS Markit PEP reports. See Supplementary Information for full list of external reports. Some process data includes integrated steps, leading to intermediate nodes being excluded, for example lignite to propylene. The final products selected are the polymers: low density polyethylene (LDPE), linear low-density polyethylene (LLDPE), bimodal high-density polyethylene (HDPE), polypropylene (PP), polyethylene terephthalate (PET), polyvinyl chloride (PVC), polybutylene terephthalate (PBT) and polyacrylic acid (PAA). Plotting a chemical network in this way allows for quick insights, crucially on where the most carbon intensive steps lie in the plastics value chain. Typically, these are the first or second chemical transformations, with the final polymerisation steps being less carbon intensive. Therefore, more investment and scientific effort should be placed on reducing the carbon intensity of the initial key chemical feedstocks such as ethylene and methanol.

This graph is limited in its use, however, as it does not account for the feedstock consumption of each step toward the targeted products. Furthermore, the above visualisation does not show every possible route (of which there are over 23,000 to the polymer end products from the over 200 processes included), instead only showing the least carbon intensive step between each feedstock and product.

Interacting with the Network

Further complexity (and more information) may be introduced by selecting specific chemical pathways via an interactive user interface, such as the one shown in Figure 5. This user interface allows the user to select a region, a target product from the network, a starting feedstock, any intermediate products to include in the pathway and any specific processes to include as well. The output is a breakdown of the metrics of every possible route in the network that fulfils the user restrictions. This can be used to quickly overview different process routes and see where the most carbon intensive steps are. Searching for process routes to a specific product, with the option to select starting feedstocks, intermediate products and specific processes, takes less than one minute. Data entry for a new process, assuming the data is available, takes less than five minutes. This compares favourably with much more detailed LCA approaches, which can take weeks or months to complete.

An example of using this network for high-level, early-stage process assessment is shown in Figure 6. Here, the effect of changing the raw material when making PP via methanol is investigated. For each process step, the lowest equivalent CO₂ process was selected, with Steps 2 and 3 identical for Figure 6(a), 6(b) and 6(c). All equivalent CO₂ values were calculated according to the methodology described. Equivalent CO₂ from other sources, such as materials sourcing, transport and leaks, are not included. The selected region is the USGC, with direct utilities supplied via natural gas, and indirect utilities sourced from the regional energy mix, except for Figure 6(c) where indirect utilities is approximated as 0 tonne_CO2e kWh^–1. Direct air capture CO₂ is assumed to have a component CO₂e of 0 tonne_CO2e tonne_CO2^–1.

From this analysis, some initial conclusions may be drawn. Firstly, the final polymerisation step has a lower net CO₂e than the propylene formation step and the methanol formation step in Figure 6(a) and 6(b). In addition, the vast majority of the polymerisation net CO₂e stems from the indirect utilities used. By using renewable electricity, this value could be drastically reduced. This is also shown in full LCA analysis of PP production, with cradle-to-propylene production accounting for 82% of the CO₂e of PP production and the majority of utilities in a PP resin plant used for electricity generation (43).

Secondly, the dramatic effect of changing the feedstock used can be seen, with the use of coal over natural gas leading to an over six-times increase in the overall carbon intensity. On the other hand, using direct air capture CO₂, green hydrogen and renewable electricity to make methanol leads to a significantly carbon negative process.

Finally, the chemical networks approach allows the user to view the ‘big picture’ effect of their first process step choices. By choosing to use a green methanol process, the carbon intensity of the propylene formation and polymerisation steps may be cancelled out, leading to overall carbon-negative PP. What’s more, captured CO₂ produced from the natural gas utilities could then be reused in the first step in an integrated plant, reducing the electricity requirements.

These conclusions could help direct resourcing efforts to the most impactful changes to reduce the CO₂e of PP and therefore reach the industry’s net zero goals.

Comparison with Literature

While a comparison with literature for every process analysed using this methodology is not feasible, Figure 7 presents a comparison for a variety of processes used to make ethylene. As shown in Figure 7, the chemical networks methodology presented in this paper provides comparable results with other data available in the literature for the ethylene production processes.

A weakness of the current methodology is highlighted in Figure 8. Applying the chemical networks methodology to estimate the CO₂ per tonne for the production of PVC from ethane in the USGC results in a substantially lower CO₂e value compared to the values estimated by LCA using data from ecoinvent^TM (1.35 tonnes_CO2e tonne_PVC^–1 vs. 2.59 tonnes_CO2e tonne_PVC^–1 respectively). However, if the indirect material sourcing equivalent CO₂ for the chlorine is included (1.09 tonnes_CO2e tonne_PVC^–1), then the results produced by the two methods are comparable (2.44 tonnes_CO2e tonne_PVC^–1 vs. 2.59 tonnes_CO2e tonne_PVC^–1). It is expected that the chemical networks result would be lower than LCA, as variables such as transport are not factored in. This indicates the importance of including this variable in further work.

Advantages of the Chemical Networks Methodology

The Chemical Networks methodology was designed to quickly assess the sustainability metrics of process routes from raw material to final product and it is successful in rapidly assessing the environmental impact of tens of thousands of pathways in minutes. Such speed is valuable for early-stage decision making for investment or policy decisions and where reliable LCA data is scarce. The ease of use of the tool also encourages users to consider the sustainability consequences of their decisions at the start of new project cycles, as timely initial assessments are no longer a barrier to do so.

What’s more, if the process data is available (or can be reasonably estimated), adding new processes and expanding the network is trivial. This is essential as new, sustainable processes (such as bio-based or electrochemical processes) are commercialised and require quick comparison with existing routes.

Further Work

While this work serves as an initial proof of concept for the methodology, further development will enhance the utility and accuracy of the tool developed. A notable weakness of the current network is its inability to meaningfully handle multiple feedstocks and products as branching paths. By creating a binodal graph, where feedstocks and products act as one node and processes as another, this could be overcome. This is illustrated in Figure 9, where factoring in multiple products is greatly simplified through the binodal approach. By containing all the process data in nodes, rather than in both nodes and edges, branching path computation is greatly simplified. This approach was undertaken by Grzybowski et al. in their research on retrosynthesis of organic reactions using a chemical networks approach (14–21). The user could interact with the network in the same manner as the current implementation but limit the length of paths unrelated to the target product to prevent the computation reaching unfeasibly expensive levels.

Another addition, discussed earlier, is the inclusion of indirect material sourcing metrics to generate more realistic CO₂e levels for different chemical feedstocks. Additional process metrics, such as cost, may be added to provide even more meaningful comparison between different chemical pathways. Adding additional variables to the network is trivial as it simply requires the addition of information to the starting database.

In their current state, the methodology and results do not meet the LCA standards of ISO 14044:2006 (47). This was out of scope for this initial proof of concept. Instead, this work is designed to complement LCA studies by providing an initial, high-level view for early-stage investigations. While meeting the LCA International Organization for Standardization (ISO) standard steps (goal and scope definition, life cycle inventory analysis, life cycle impact assessment and interpretation) for each pathway is not feasible, the methodology could be adapted and used with established LCA software, data and methodologies to meet these standards while maintaining the intuitive design.

Finally, the chemical network may be vastly expanded to include a further range of chemical industries. While this work has focused on the plastics industry, through the simple methodology provided and with the process data available, adding additional processes and routes to other chemicals is trivial. As the number of processes used in large-scale industrial chemical synthesis is rather more limited than, for example, the pharmaceutical industry, providing a meaningful picture of this industry is not implausible.

Conclusion

To meet the Paris Agreement pledge and net zero targets, the chemical industry requires data-driven analysis methods for early-stage decision making, to better target resources and accelerate delivery on sustainability commitments. Current LCA methods, though acting as the gold-standard for data accuracy and transparency, are limited for this application by their necessary high labour intensity (taking weeks or months to complete) and their site or process specificity.

In this work, the sustainability metrics (equivalent CO₂, water consumption, feedstock consumption, E-factor) of over 23,000 chemical pathways in the plastics industry were investigated using graph theory principles. The resulting chemical network allows for rapid (~1 min) insights for early-stage sustainability assessments, saving considerable time compared to more in-depth LCA methods and indicating promising low-carbon routes or areas where more investment and innovation into low carbon solutions is needed.

The analysis produced reasonable results when compared to literature, but further work on including indirect, or Scope 3, emissions and adapting to LCA ISO standards is required to improve the accuracy and quality. The interactive tool, while accessible to a wide range of users, would benefit from branching paths for multi-product processes, which could be achieved via a binodal graph structure. Further data such as costs may also be added in future development.

The chemical networks tool has great future potential, with the design allowing new processes to be quickly and easily added. The scope could therefore feasibly be expanded to include nearly all large-scale industrial chemical processes, allowing for a high-level overview of the entire industry. As the need for drastic reductions in emissions to reach net zero targets grows, such insight may prove invaluable.