Industries / Mining and materials / Chemicals and petrochemicals / Specialty chemicals
Batch production
Batch production makes a product one vessel-load at a time, following a recipe of charges, temperatures and hold times, often in multipurpose plants where the same vessels make many products in turn.
5
Milestones achieved, from 1884 to 2017.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1884
Glass-lined steel tanks
Caspar Pfaudler, a German brewmaster in Rochester, New York, invented Glasteel, steel lined with glass, according to the company that still makes it. The company says the growing chemical industry took it up by 1920 because glass resists corrosion. Glass-lined vessels are still common in batch plants: in 2006 the contract manufacturer Lonza reported that around half the equipment in its new multipurpose plant was glass-lined, because corrosive materials are often handled there.
Credit
Caspar Pfaudler
1997
A standard for batch recipes
The International Electrotechnical Commission (IEC) published IEC 61512-1, the international form of the US batch control standard ISA-88. It gave batch plants shared models and words for recipes, equipment and procedures, including a "general recipe" that describes a product's process without naming any particular equipment. Free XML schemas for exchanging such recipes, called BatchML, are published by MESA International, and the IEC replaced the 1997 text with a new edition in 2026.
Credit
the International Society of Automation (ISA) and the IEC
1998
A dye batch runs away
On 8 April 1998 a 2,000-gallon kettle making Automate Yellow 96, a dye used to tint petroleum fuels, overheated at Morton International's plant in Paterson, New Jersey. The reaction ran away and blew the hatch off the kettle, and nine workers were injured. The US Chemical Safety Board (CSB) found that in 1990 the plant had switched from adding one raw material in four equal portions to adding all of it at the start, that a 1996 move to a bigger kettle and batch made cooling harder, and that eight earlier temperature overruns had never been investigated.
LessonFeeding a reactive ingredient in portions limits how much can react at once, and every temperature overrun is a warning to investigate.
Credit
US Chemical Safety Board
2006
A recipe changed without review
On 31 January 2006 a runaway reaction and vapor cloud explosion destroyed Synthron's plant in Morganton, North Carolina, which made powder coating and paint additives by polymerizing acrylic monomers in a 1,500-gallon reactor. One worker was killed and 14 were injured. To fill an order slightly larger than usual, managers scaled up the recipe and put all the extra monomer into the first charge; the CSB says this more than doubled the rate of heat release, beyond what the reactor's condenser could remove, and that recipes were changed without systematic review.
LessonEven a small change to a batch recipe needs the hazard review a new process would get.
Credit
US Chemical Safety Board
2017
Floodwater beats every backup
Arkema's plant in Crosby, Texas makes organic peroxides, which start the reactions that make many plastics and cure rubber, and which must be kept cold. Hurricane Harvey's flooding knocked out its power, backup generators, liquid-nitrogen cooling and refrigerated trailers, and on 31 August 2017 the peroxides began to decompose and burn. Over three fires more than 350,000 pounds burned, 21 people sought medical attention, and more than 200 residents could not return home for a week. The CSB found that every layer of protection failed for the same reason, flooding.
LessonBackups that share one weakness, such as standing on the same floodable ground, act as a single safeguard.
Credit
US Chemical Safety Board
Watch for
Fluorine intermediate made in flow
Anupam Rasayan India says it is the first company in the world to make ethyl trifluoroacetate (ETFA), a building block for fluorinated drugs, crop chemicals and materials, by continuous flow chemistry at industrial scale. In flow chemistry the ingredients react while moving through a pipe or channel, instead of sitting together in one large vessel. The report gives no capacity or output.
Achieved whenoutput from the flow line is confirmed by an outside party, such as a customer, an auditor or a regulatory filing.
Claimed by
Anupam Rasayan India, June 2026
Viscosity read by camera
Researchers at Seoul National University report ViscNet, a video model that estimates a stirred liquid's viscosity (its thickness) from how a patterned background is distorted by the moving surface. On water–glycerin mixtures sorted into ten viscosity classes it named the right class 80.7% of the time, and its regression error was 0.113 in log units.
Achieved whenan outside lab or a manufacturer reports similar accuracy on its own vessels and liquids.
Claimed by
Jongwon Sohn, Juhyeon Moon, Jaewook Nam and colleagues, Seoul National University, December 2025
Open
01
Procedures into checked batch recipes
Software that reads a written batch procedure, such as a worked example in a patent, and writes it as a structured recipe in BatchML, with each charge, heating step, hold and transfer as its own step. A second program checks the recipe against plain rules a reviewer applies: a reactive ingredient is fed in portions or at a set rate rather than all at once, every heating step has a maximum temperature and a cooling action, every hold has a time. It then lists what the written text leaves out.
The briefCredit · unclaimed02
Camera viscosity for opaque batches
A camera and model that estimate the viscosity of a stirred liquid that is opaque and shear-thinning (it gets thinner the faster it is stirred, as paint does) from video of its surface while the stirrer keeps running, and that follow a batch as it thickens, so the end point can be called without pulling samples.
The briefCredit · unclaimed
Open 01 · the brief
Procedures into checked batch recipes
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Software that reads a written batch procedure, such as a worked example in a patent, and writes it as a structured recipe in BatchML, with each charge, heating step, hold and transfer as its own step. A second program checks the recipe against plain rules a reviewer applies: a reactive ingredient is fed in portions or at a set rate rather than all at once, every heating step has a maximum temperature and a cooling action, every hold has a time. It then lists what the written text leaves out.
Why it is still open
IBM Research's Paragraph2Actions (2020) turns written lab procedures into sequences of lab actions, and matched a person's sequence exactly for 60.8% of sentences, but it writes no plant recipe and checks no hazards. LLM4SFC (December 2025) turns plain-language descriptions of factory workflows into controller programs and reports 75–94% generation success, which measures whether the program is valid, not whether it matches the procedure, and it was tested on private data. RWTH Aachen's ModPlant-LLM (2026) plans plant operations from a general recipe that is already in XML, not from written text. Lee Cronin's group compiled 12 published procedures into code for a lab robot in 2020, not into plant recipes. A search found no public set of written procedures paired with checked batch recipes.
Built with what exists
The BatchML schemas, free from MESA International on GitHub; RWTH Aachen's open-source RecipeEditor, which draws and validates BatchML recipes; IBM's open-source Paragraph2Actions as a baseline for splitting text into steps; worked examples from patents, which are free to read; an open-weight model to read the text, such as Mistral Large 4 (see the Note "A trillion open parameters"); a coding agent that tests its own output, such as one built on Inkling 2 Small (see the Note "Checked without being asked"). ISA forbids entering its standards into AI tools, so the rules should be written from the public BatchML schema and published guidance, not by pasting the standard.
First step
Pick 10 patent examples for coating or adhesive additives and draw their recipes by hand in RecipeEditor to settle the format and the rule list.
Done looks like
100 patent examples converted. Two chemical engineers, working separately, write recipes for 30 of them, and the software matches their steps (order, amounts and temperature limits) on at least 90% of steps. In 20 further examples with omissions planted on purpose, the checker finds at least 18. Code, recipes and scores are public.
Open 02 · the brief
Camera viscosity for opaque batches
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A camera and model that estimate the viscosity of a stirred liquid that is opaque and shear-thinning (it gets thinner the faster it is stirred, as paint does) from video of its surface while the stirrer keeps running, and that follow a batch as it thickens, so the end point can be called without pulling samples.
Why it is still open
ViscNet's authors write that their model is "restricted to transparent Newtonian fluids", needs pauses in mixing and "is not yet fully in-line"; its real test liquids were water–glycerin mixtures up to about 130 times as viscous as water. A 2026 conference abstract on making monomers for adhesives says viscosity is traditionally measured only by offline lab tests, so problems are found too late.
Built with what exists
Harmless test liquids: water thickened with xanthan gum or cornstarch, made opaque with milk powder; a rotational viscometer for reference readings; a lab overhead stirrer and a phone or Raspberry Pi camera; SPlisHSPlasH, the open-source particle fluid simulator ViscNet used to make training videos; a video model of the kind ViscNet uses, a video vision transformer.
First step
Film 20 one-minute videos of four xanthan gum strengths in a 2-liter beaker at three stirrer speeds, and record a viscometer reading for each.
Done looks like
On 12 liquids, at least 6 of them opaque and shear-thinning, the camera estimate is within 20% of the viscometer reading at a fixed speed on at least 80% of 200 held-out videos, with the stirrer running. On 10 thickening runs (xanthan gum hydrating in water), it calls the end point within 2 minutes of the viscometer. Videos, readings and code are public.