Industries / Mining and materials / Chemicals and petrochemicals / Refining
Cleaning and blending fuels
Treating removes sulfur and other unwanted parts of each fuel stream, often by reacting it with hydrogen over a catalyst, and blending mixes the treated streams into gasoline (petrol), diesel and jet fuel that meet the legal specification.
5
Milestones achieved, from 1883 to 2020.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1883
Sulfur recovered from gas
Carl Friedrich Claus was granted a patent in late 1883 for a process that turns hydrogen sulfide, a toxic gas, into elemental sulfur. In 1938 the German company I.G. Farbenindustrie changed it significantly, giving the modified Claus process. A 1986 review by B. G. Goar estimated that it made 90 to 95% of the world's recovered sulfur, in over 380 plants, and that most sour crude oil, meaning crude high in sulfur, ends up sending its hydrogen sulfide to it. The US EPA puts recovery at 94 to 96% for a two-stage plant and says the reaction itself stops at 97 to 98%; tail-gas units added after the plant raise recovery to 98 to 99%, and some scrubbers to about 99.9%.
Credit
Carl Friedrich Claus; I.G. Farbenindustrie
1926
A scale for engine knock
Knock is the metallic rattle of fuel in an engine cylinder exploding on its own instead of burning smoothly, and nobody could tell in advance whether a gasoline would cause it. Graham Edgar of the Ethyl Gasoline Corporation, then part of General Motors and Standard Oil, found that isooctane did not knock in any engine and n-heptane always did. A fuel's octane number became the percentage of isooctane in the mix of the two that matched it in a test engine. Commercial gasoline then rated 40 to 75. Edgar published the method in January 1927.
Credit
Graham Edgar, Ethyl Gasoline Corporation
2006
An additive withdrawn
MTBE (methyl tertiary-butyl ether) was blended into US gasoline from 1979 to replace lead as an octane booster. From 1992 it was used at higher levels to meet the oxygen content that the 1990 Clean Air Act Amendments required in some cities, and US output passed 200,000 barrels a day in 1999. MTBE dissolves easily in water. Studies found it in groundwater across the country, and the EPA says low levels can make drinking water undrinkable through taste and odor. The Energy Policy Act of 2005 removed the oxygen requirement for reformulated gasoline, which ended in May 2006, and refiners switched from MTBE to ethanol.
LessonA component chosen for its effect on air was not judged on what it did to water when fuel leaked.
Credit
the US Congress, by the EPA's account
2006
Sulfur cut to traces
Before regulation, US diesel held as much as 5,000 parts per million (ppm) of sulfur, by the EPA's account. The EPA began limiting it in 1993, and from 2006 to 2010 phased in ultra-low-sulfur diesel, at 15 ppm, for road vehicles. In Europe a 2003 directive required gasoline and diesel with at most 10 mg/kg of sulfur to be on sale from 2005 and to be the only kind sold from 1 January 2009.
LessonCleaner fuel at the pump can mean more carbon dioxide at the refinery, and the EU set its dates so that savings from new vehicles would outweigh it.
Credit
the US EPA; the European Union
2020
India skips a fuel grade
On 1 April 2020 all of India's more than 68,700 fuel stations sold BS-VI gasoline and diesel, the equivalent of Europe's Euro VI grade, with 10 ppm of sulfur, Indian Oil's chairman told the Press Trust of India. The country had moved to BS-III fuel at 350 ppm in 2010 and to BS-IV at 50 ppm seven years later, and then went straight to BS-VI in three years, skipping BS-V. State-owned refiners spent about Rs 35,000 crore on the upgrade, on top of Rs 60,000 crore for the earlier switches, by the same report.
LessonGoing straight to the cleaner grade meant refiners and carmakers invested once instead of twice, by Indian Oil's account.
Credit
the Government of India; Indian Oil, Bharat Petroleum, Hindustan Petroleum and the other refiners
Watch for
Hydrogen from sulfur gas
A team at the Dalian Institute of Chemical Physics led by Li Can says its electrochemical method splits hydrogen sulfide into hydrogen and sulfur at ambient conditions; China Daily contrasts this with the Claus process, which needs heat above 1,200 °C. A pilot in Xinxiang, Henan province, treating waste gas from a methanol plant, ran more than 1,000 hours without a stop, by the account of the general manager of the partner company, Shandong Sunway Chemical Group. It made sulfur above 99.95% pure and hydrogen 99.999% pure. Experts at an evaluation meeting in Beijing recommended scale-up. No figure for the electricity used was given.
Achieved whena unit runs on a refinery's hydrogen sulfide and the refiner or an outside party publishes the share converted, the hours run and the electricity used per kilogram of hydrogen.
Claimed by
Li Can's team and Shandong Sunway Chemical Group, January 2026
Open
01
Sulfur plants, refinery by refinery
A public table that pairs each US refinery's sulfur recovery capacity with the sulfur dioxide (SO2) its sulfur plant reports to the US EPA, works out SO2 per short ton of sulfur, and flags plants far above what EPA figures say plants with tail-gas units achieve.
The briefCredit · unclaimed02
A test for blend-planning AI
A public set of gasoline and diesel blending problems, each with its components, their qualities and prices, a legal specification and a solver-checked cheapest recipe, used to score AI models that are asked to plan the blend.
The briefCredit · unclaimed
Open 01 · the brief
Sulfur plants, refinery by refinery
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table that pairs each US refinery's sulfur recovery capacity with the sulfur dioxide (SO2) its sulfur plant reports to the US EPA, works out SO2 per short ton of sulfur, and flags plants far above what EPA figures say plants with tail-gas units achieve.
Why it is still open
The EPA's AP-42 emission factors run from 57 lb of SO2 per ton of sulfur for a two-stage plant with controls to 278 lb for an uncontrolled one-stage plant, and test data for one plant type put recovery anywhere from 95 to 99.8%. A search found no published plant-by-plant comparison for US refineries.
Built with what exists
The EIA's refinery capacity file, which lists each refinery's sulfur capacity in short tons a day and its desulfurization capacity by product; the point-source files of the EPA's 2020 National Emissions Inventory; Python with pandas; a language model that reads state permits for the outliers and quotes the page each figure came from (see the Note "Findings that cite notes").
First step
Download the EIA file and the inventory's point-source file, confirm that sulfur recovery units appear as separate processes, and match ten Texas refineries by hand.
Done looks like
Every refinery with sulfur capacity in the EIA file matched to its emission records; SO2 per short ton of sulfur capacity published for each; the ten highest checked by a second person against state permit or emission-report documents; code and table public.
Open 02 · the brief
A test for blend-planning AI
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of gasoline and diesel blending problems, each with its components, their qualities and prices, a legal specification and a solver-checked cheapest recipe, used to score AI models that are asked to plan the blend.
Why it is still open
Blending has been treated as a linear program since 1952, and open blend optimizers are common on GitHub. General tests of AI models on optimization problems exist, such as NL4Opt and IndustryOR, which its authors call the first industrial benchmark of its kind, and in February 2026 MIPLIB-NL found that systems that do well on those tests do much worse on real industrial models. None is built on fuel blending with legal specifications. The nearest test found, a 2026 student project, scored answers to 76 questions about a blending procedure document.
Built with what exists
HiGHS, an open-source solver, or Pyomo, to compute each answer; legal limits such as the EU's 10 mg/kg sulfur cap; several AI models, including an open-weights one such as Mistral Large 4 (see the Note "A trillion open parameters"); each problem run 20 times per model, since a model that passes once may fail the next time (see the Note "All twenty times").
First step
Write 20 small problems with four to six components each and solve them with HiGHS.
Done looks like
300 problems with answer keys under an open license; at least three models scored; for each model, the share of problems it solves within the specification and within 0.5% of the cheapest cost on all 20 runs.