TL;DR: A team led by Bhavik R. Bakshi reported on September 16, 2026 that pairing large language models with retrosynthesis built the broadest circular network for methanol, with 165 reactions among 12 chemicals, and scaled to more than 1,000 reactions for polyethylene.

A life-cycle route map instead of a gate-to-gate recipe

The central result, published online on September 16, 2026 in ACS Sustainable Chemistry & Engineering, is that manufacturing routes that explicitly account for the entire product life cycle can be explored systematically when language-based knowledge extraction is paired with structure-informed retrosynthetic reasoning.

According to ACS Sustainable Chemistry & Engineering, discovery and development of manufacturing routes that explicitly account for the entire product life cycle are essential for the transformation to a sustainable and circular chemical industry.

The same source describes retrosynthesis as a promising approach that operates within a gate-to-gate paradigm, a framing that limits its ability to explicitly integrate end-of-life waste streams into upstream production pathways.

The authors therefore evaluated four reaction discovery strategies that combine human intervention, pattern recognition, retrosynthesis, and large language models within a common comparative framework, according to ACS Sustainable Chemistry & Engineering.

The paper is identified by ACS Sustainable Chemistry & Engineering as a Research Article dated September 16, 2026, with publication in issue September 28, 2026 in volume 14, issue 38, pages 17235–17251.

Who reported the work and when it cleared review

The listed authors are Sunghoon Kim, Avan Kumar, Hariprasad Kodamana, Manojkumar Ramteke, and Bhavik R. Bakshi, with Bhavik R. Bakshi identified as the corresponding author, according to ACS Sustainable Chemistry & Engineering.

The affiliation information carried by ACS Sustainable Chemistry & Engineering associates Sunghoon Kim, Avan Kumar, and Bhavik R. Bakshi with Arizona State University and Hariprasad Kodamana and Manojkumar Ramteke with the Indian Institute of Technology Delhi.

Contact information published with the article lists [email protected] for the corresponding author, as recorded by ACS Sustainable Chemistry & Engineering.

The publisher is listed as the American Chemical Society, according to ACS Sustainable Chemistry & Engineering.

The timeline published with the article states Received June 07, 2026, Revision Received September 01, 2026, Accepted September 01, 2026, and Published Online September 16, 2026, according to ACS Sustainable Chemistry & Engineering.

The funding information published with the article names the Ira A. Fulton Schools of Engineering at Arizona State University and the Science and Engineering Research Board, including award identifier VJR/2018/000123, according to ACS Sustainable Chemistry & Engineering.

The competing-interest statement published with the article states that the authors declare no competing financial interest, according to ACS Sustainable Chemistry & Engineering.

Why end-of-life streams were the missing link

The problem definition in the article centers on integration of end-of-life waste streams into upstream production pathways, a capability that gate-to-gate retrosynthesis alone does not explicitly provide, according to ACS Sustainable Chemistry & Engineering.

That distinction matters for the proposed circular chemical reaction network, abbreviated CCRN, which is intended to support systematic exploration of candidate circular chemical pathways for subsequent economic and environmental evaluation, according to ACS Sustainable Chemistry & Engineering.

The supporting-information description indicates that the comparison included algorithms for reaction extraction, compound screening, and stoichiometric reconciliation, according to ACS Sustainable Chemistry & Engineering.

The same supporting-information description lists prompt-response records and model comparisons, methanol and polyethylene compound and reaction datasets, and TRL classifications with supporting references, according to ACS Sustainable Chemistry & Engineering.

It further lists expanded reaction-network diagrams, evaluation and sensitivity-analysis results, and end-of-life reaction-source information, according to ACS Sustainable Chemistry & Engineering.

The supporting information is described as available free of charge at the article DOI, according to ACS Sustainable Chemistry & Engineering.

Four strategies tested on one common framework

The comparative design is a key element of the September 16, 2026 report, because all four reaction discovery strategies were assessed within a common comparative framework, according to ACS Sustainable Chemistry & Engineering.

Those four inputs are identified as human intervention, pattern recognition, retrosynthesis, and large language models, described as LLM in the article record, according to ACS Sustainable Chemistry & Engineering.

A listing carried by ScholarBro for the same September 2026 paper in ACS Sustainable Chemistry & Engineering likewise records five authors and 44 references, with the same title and journal.

The ScholarBro listing repeats the finding that language-based knowledge extraction combined with structure-informed retrosynthetic reasoning enables scalable construction of CCRNs, consistent with the abstract carried by ACS Sustainable Chemistry & Engineering.

Methanol benchmark: breadth versus maturity

When evaluated with methanol as a benchmark system, the strategies reveal distinct trade-offs between pathway discovery and technological maturity, according to ACS Sustainable Chemistry & Engineering.

Combining LLMs with retrosynthesis achieves the broadest expansion of reaction space in that methanol comparison, according to ACS Sustainable Chemistry & Engineering.

The reported scale for that combined approach is 165 reactions involving 12 unique chemicals, according to ACS Sustainable Chemistry & Engineering.

The same combined approach is reported to have the highest novelty relative to the conventional business-as-usual reference network, according to ACS Sustainable Chemistry & Engineering.

In other words, the methanol test did not present one method as superior on every dimension; it presented a trade-off in which broader discovery and higher novelty were balanced against technological maturity, according to ACS Sustainable Chemistry & Engineering.

The reference point for that novelty statement is explicitly the conventional business-as-usual reference network, according to ACS Sustainable Chemistry & Engineering.

From 12 chemicals to more than 1,000 reactions

For a more complex molecule, polyethylene, the resulting circular chemical reaction network contains more than 1,000 reactions, according to ACS Sustainable Chemistry & Engineering.

That polyethylene result is presented as the demonstration that the framework can scale beyond the 12-chemical methanol benchmark system, according to ACS Sustainable Chemistry & Engineering.

Human-guided extraction increased the total number of identified reactions by 34–35%, depending on the keyword-search strategy, according to ACS Sustainable Chemistry & Engineering.

The same polyethylene finding states that human-guided extraction recovered additional end-of-life reactions from experimental results, according to ACS Sustainable Chemistry & Engineering.

The dependence on keyword-search strategy is therefore part of the reported outcome, not a background detail, because the 34–35% range is tied to the strategy used, according to ACS Sustainable Chemistry & Engineering.

What the proposed framework does and does not claim

The proposed framework demonstrates that combining language-based knowledge extraction with structure-informed retrosynthetic reasoning enables scalable construction of CCRNs, according to ACS Sustainable Chemistry & Engineering.

It is also described as supporting the systematic exploration of candidate circular chemical pathways for subsequent economic and environmental evaluation, according to ACS Sustainable Chemistry & Engineering.

That wording positions economic and environmental evaluation as a subsequent step for candidate pathways, rather than as a completed assessment inside the network-construction report, according to ACS Sustainable Chemistry & Engineering.

The emphasis on candidate pathways likewise indicates that the networks support exploration of options, rather than selection of a single manufacturing route, according to ACS Sustainable Chemistry & Engineering.

How this fits wider retrosynthesis work published that week

Retrosynthetic analysis continues to focus on identifying strategic bonds whose disconnection simplifies a target molecule into accessible precursors, a challenge described in a September 14, 2026 Perspective in The Journal of Organic Chemistry titled Line-Graph Node Centrality in Retrosynthesis.

That Perspective states that computational approaches to synthesis planning have evolved from rule-based systems to artificial intelligence platforms, while noting a need for interpretable, quantitative, structure-based metrics that guide bond disconnection.

The September 16, 2026 circular-network paper approaches the same planning challenge from the life-cycle direction, asking how upstream production pathways can explicitly include end-of-life streams, according to ACS Sustainable Chemistry & Engineering.

The Perspective in The Journal of Organic Chemistry demonstrates its metric on six structurally diverse natural products named as longifolene, hetisine-2,11,13-triol, auriculatol A, ineleganolide, quinocarcin, and papililone A, followed by seven additional examples.

That context helps explain why the circular-network authors paired structure-informed retrosynthetic reasoning with language-based extraction, rather than relying on either capability alone, according to ACS Sustainable Chemistry & Engineering.

Maturity remains part of the comparison

A separate September 16, 2026 article in the Journal of Chemical Information and Modeling, titled The Synthesizability Gap Has Moved: Route Existence Is Not Chemistry, argues that synthesis-constrained generative models now guarantee that every proposed molecule carries a route to purchasable building blocks.

That article states that recent campaigns have carried such designs to synthesized, biologically active compounds, while arguing that a route a planner certifies is not yet chemistry a bench reproduces.

The relevance to the circular-network study is the shared emphasis on maturity: the methanol benchmark explicitly reports trade-offs between pathway discovery and technological maturity, according to ACS Sustainable Chemistry & Engineering.

The supporting-information record for the circular-network paper reinforces that emphasis by listing TRL classifications and supporting references among the deposited materials, according to ACS Sustainable Chemistry & Engineering.

What to watch in the deposited datasets and diagrams

The deposited record for the circular-network study includes methanol and polyethylene compound and reaction datasets, according to ACS Sustainable Chemistry & Engineering.

It also includes TRL classifications and supporting references, expanded reaction-network diagrams, and evaluation and sensitivity-analysis results, according to ACS Sustainable Chemistry & Engineering.

The inclusion of algorithms for reaction extraction, compound screening, and stoichiometric reconciliation indicates that network construction required reconciliation and screening steps in addition to extraction, according to ACS Sustainable Chemistry & Engineering.

The inclusion of prompt-response records and model comparisons indicates that language-model behavior was documented alongside the resulting networks, according to ACS Sustainable Chemistry & Engineering.

The inclusion of end-of-life reaction-source information indicates that the source of waste-stream reactions was tracked separately, according to ACS Sustainable Chemistry & Engineering.

For readers tracing reproducibility, the article history — received June 07, 2026, revised September 01, 2026, accepted September 01, 2026, and published online September 16, 2026 — provides the complete editorial timeline published by ACS Sustainable Chemistry & Engineering.

The issue assignment — September 28, 2026, volume 14, issue 38, pages 17235–17251 — provides the citable location published by ACS Sustainable Chemistry & Engineering.

Together, the methanol benchmark at 165 reactions involving 12 unique chemicals and the polyethylene network at more than 1,000 reactions define the two reported scale points for the framework that combines language-based knowledge extraction with structure-informed retrosynthetic reasoning, according to ACS Sustainable Chemistry & Engineering.

Sources referenced