The National Laboratory of the Rockies (NLR) is developing three super intelligence systems that could help researchers predict biological behavior, identify industrial processing problems and design proteins for commercial applications. The projects target a persistent obstacle in biotechnology: scientific discoveries made in laboratories do not always translate into reliable performance at industrial scale. By combining historical experiments, machine learning and computer-generated scientific data, NLR researchers want to give companies better predictions before they commit substantial resources to new processes.
The initiative builds on more than 45 years of biological and bioprocessing research at the US Department of Energy laboratory. Its bioenergy and bioeconomy program has generated thousands of scientific publications and accumulated extensive experimental records. According to NLR's November 11 report, researchers are now organizing that knowledge into systems designed to make scientific insights more accessible, identify relationships across different datasets and help industry make better-informed technical decisions.
A large scientific data resource
NLR's bioenergy and bioeconomy program has accumulated a substantial research record since 1979. Its work has produced more than 3,600 peer-reviewed publications, reports and book chapters, covering subjects that range from biomass characterization and thermochemical processing to synthetic biology and the challenges of increasing production from laboratory experiments to commercial operations.
These publications contain more than descriptions of scientific discoveries. They also document measurements collected during experiments. Depending on the research, those records can include product yields, material balances, gas released during processing, acidity levels and other parameters that help researchers understand how biological and chemical systems behave.
The scale and variety of this historical information provide an opportunity for computational research. However, turning decades of scientific records into a useful predictive resource requires more than collecting documents and feeding them into a computer model. Researchers must understand which measurements are available, how consistently experiments were conducted and whether the information is sufficiently complete to support reliable conclusions.
That challenge has become particularly important in 2026, following the launch of the Department of Energy's Genesis Mission and advances in fundamental super intelligence models. The mission and developments in computational capabilities have increased the importance of making scientific information accessible in formats that advanced computational systems can use.
Commercial chatbots demonstrate how large language models can learn patterns from enormous collections of written material and generate relevant responses. Scientific applications require a more specialized approach. Models need access to carefully selected experimental measurements, relevant scientific publications and instructions that help them retrieve information, use supporting computational tools and present results in a useful format.
NLR researchers are applying that approach to scientific workflows. Rather than treating a model as a general-purpose source of answers, they want to connect it to specialized datasets and computational methods that can help researchers complete technical tasks, explore possible outcomes and identify relationships that might otherwise remain buried in years of research.
Uneven data creates opportunities
One of the biggest obstacles is that biological research does not produce equally comprehensive datasets across every field. Some research areas already have large collections of consistent measurements, while others lack sufficient information to train and validate predictive systems directly.
Juliane Mueller, an applied mathematician at NLR, identifies protein engineering as an area where researchers can draw on thousands of structured and relatively consistent data points. Although those measurements may come from experiments conducted at comparatively small scales, they can provide enough information to develop and evaluate predictive models.
The situation differs considerably in pilot-scale and industrial-scale bioprocessing. These operations involve larger equipment, more complicated operating conditions and interactions that may not appear in smaller laboratory experiments. High-quality datasets covering these environments remain limited, making direct prediction more difficult.
Mueller argues that this imbalance does not prevent super intelligence systems from contributing to research. Instead, researchers can combine reliable measurements from data-rich fields with additional methods for addressing gaps in less thoroughly documented areas.
Those methods include synthetic data and multifidelity approaches. Synthetic data consists of computer-generated information designed to represent scientific behavior under defined assumptions. Multifidelity approaches combine information produced through methods with different levels of detail, accuracy or computational cost.
The distinction matters because generated data cannot simply be treated as equivalent to experimental observations. Researchers need to understand how it was produced, which assumptions shaped it and where its predictions are trustworthy. When used responsibly, however, synthetic information can expand the range of scenarios that researchers can examine.
"These tools offer a way to make the most of a mixed data landscape."
— Juliane Mueller, NLR applied mathematician
The laboratory's three projects apply these principles to different problems. One focuses on identifying failures before biorefineries reach commercial scale. Another examines how microorganisms behave inside large fermentation equipment. The third uses protein models and experimental measurements to identify biological molecules with useful industrial characteristics.
Predicting biorefinery failures
The first application addresses the difficulties companies encounter when transforming agricultural waste into commercially valuable products. The United States has considerable quantities of agricultural residue that could serve as raw material for bio-based fuels, chemicals and materials.
According to the Department of Energy's national biomass resource assessment, known as the Billion-Ton report, the country could potentially collect as much as 175 million tons of agricultural residue annually. This material includes fibrous agricultural leftovers that could be processed into products beyond conventional food and feed applications.
One potential route involves producing second-generation ethanol. Unlike conventional ethanol facilities that primarily process sugars from corn kernels, these refineries are designed to convert cellulose-rich agricultural material into alcohol. Cellulose is a structural component of plant matter, and breaking it down into fermentable sugars requires additional processing steps.
Commercializing these technologies has proved difficult. Zia Abdullah, a laboratory program manager at NLR, explains that several cellulosic biorefineries have encountered technical problems when moving from individual processes to integrated commercial operations.
The difficulty is not necessarily that a particular laboratory technology fails to work. A process that performs successfully in a small experiment can encounter different pressures, material-handling requirements and operating conditions when expanded. Integrating multiple operations introduces additional interactions that researchers may not fully understand at the beginning of a project.
Previous commercial difficulties have nevertheless produced valuable information. NLR wants to combine those historical lessons with laboratory and pilot-scale measurements to train a reasoning system that can help companies recognize potential problems before investing millions of dollars in new facilities.
The project involves collaboration with Idaho National Laboratory (INL) and Lawrence Berkeley National Laboratory (LBNL), two other US Department of Energy research institutions. Together, the organizations are developing a super intelligence tool intended to identify possible failure mechanisms in proposed biorefinery processes.
The proposed system will draw on several types of evidence rather than relying on one experimental dataset. Its training material will include measurements from laboratory experiments, pilot facilities and available commercial operations. Researchers will also incorporate scientific literature, expert knowledge and synthetic information generated using mechanistic models based on fundamental physical and chemical principles.
The amount of available information varies considerably across individual processes. Some components may have only tens of relevant measurements, while others can draw on thousands of data points. The researchers therefore intend to combine evidence from different sources instead of assuming that every part of a proposed industrial system has an equally complete experimental history.
Modeling every processing step
The first technical stage involves constructing data-driven surrogate models for the individual operations needed to convert corn stover into alcohol. Corn stover refers to the stalks, leaves and other plant material remaining after corn has been harvested.
A surrogate model is a computational representation of a more complicated process. It can approximate how that process behaves under different conditions, allowing researchers to investigate scenarios without repeating every physical experiment or performing every detailed calculation from scratch.
The team is drawing heavily on information from facilities operated by its research partners. These include INL's Biomass Feedstock National User Facility, LBNL's Advanced BioProcess Development Unit and NLR's specialized laboratories and biochemical pilot plant within its Integrated Biorefinery Research Facility.
These facilities contribute experimental information from different stages of biomass processing. However, measurements for the particular applications under investigation remain limited at both laboratory and pilot scales. NLR is therefore supplementing the available observations with synthetic data generated through first-principles mechanistic models.
First-principles modeling uses established physical and chemical relationships to calculate how a system should behave. Instead of learning exclusively from recorded examples, researchers can use these models to generate additional scenarios consistent with the underlying scientific principles.
The next challenge is connecting the individual process models into a coordinated decision-support system. Mueller and her colleagues are working with Abdullah's project team to develop a tool that can help engineers examine an industrial process as a connected system rather than a collection of isolated operations.
An engineer could describe a proposed biorefinery design or introduce a particular operating scenario. The super intelligence agent would then retrieve relevant technical information from its training and supporting datasets, examine the different processing stages and provide qualitative and quantitative guidance.
Its analysis would cover several major operations involved in producing alcohol from biomass. These include preparing the feedstock, treating it with acid, breaking down material using enzymes and fermenting the resulting sugars into final products.
Each operation has its own technical requirements, but decisions made in one part of the facility can affect what happens elsewhere. A system that can reason across these connected stages could help engineers explore how individual design choices influence overall process performance.
NLR's Kestrel supercomputer, which has a computing capacity of 56 petaflops, is supporting the generation of synthetic data. A petaflop represents one quadrillion floating-point operations per second, a measure of computational performance. Kestrel provides the processing resources needed for demanding modeling and analysis tasks.
The researchers envision engineers using the system to examine practical design questions. These could include whether pressure becomes excessive near the bottom of a storage tank, what changes when a process is expanded by a factor of 1,000, or how production might be affected if biomass has remained exposed to rain for six months.
Other scenarios involve increasing process intensity to raise a facility's output. The purpose is to let engineers investigate operating conditions that could affect performance before they make expensive decisions about equipment and facility design.
A target of 2029
The biorefinery prediction project is being developed with support from the Department of Energy's Alternative Fuels and Feedstocks Office, which supports research into alternative transportation fuels and the materials used to produce them.
The current development schedule targets 2029 for completion of the model. That milestone does not automatically mean that companies will immediately be able to use it for commercial decisions. The system must first be completed and validated, a process that will establish how well its predictions correspond to the relevant scientific and engineering evidence.
Abdullah expects that, following successful completion and validation, the agent could become available for licensing to industrial companies and other interested organizations. Licensing would provide a route for external users to access the technology beyond the laboratory's internal research activities.
The proposed industrial benefit is straightforward: companies could gain an additional way to investigate potential technical weaknesses before spending millions of dollars scaling up a process. The system is being designed to support engineering decisions, not to replace the experiments and technical expertise needed to establish whether a process will work.
Matching microbes to reactors
The second application tackles a different scale-up problem. Existing ethanol facilities generally use large steel tanks containing microorganisms that ferment sugars into alcohol. These industrial vessels, known as bioreactors, provide controlled environments for biological production.
Across the United States, approximately 200 biorefineries are operating, and most produce ethanol. However, companies are increasingly interested in using similar infrastructure to manufacture a wider range of chemicals rather than limiting production to fuel alcohol.
Potential products include ethylene, an important industrial chemical used in plastics, and muconic acid, which can serve as a building block for nylon production. Expanding the range of products could allow existing facilities to explore additional markets, provided their equipment and biological processes can support the new applications.
Laboratory studies have already demonstrated that different microorganisms can convert sugar into a variety of useful products. Research published in Joule has described the potential for different microbes to produce a broad selection of chemicals through fermentation.
However, successful results in small laboratory reactors do not guarantee comparable productivity in much larger industrial vessels. Davinia Salvachúa Rodríguez, an NLR fermentation researcher and group manager, has observed cases where performance measured at bench scale does not survive the transition to larger equipment.
When productivity declines, the economics of the production process can deteriorate. A company may have selected a promising microorganism and demonstrated that it can manufacture a desired chemical, yet still struggle to achieve the output required for an economically viable industrial operation.
One possible explanation involves differences in the environment inside large tanks. Mixing does not necessarily create identical conditions throughout a vessel. Instead, different regions can develop localized variations in acidity, pressure and oxygen availability.
Researchers refer to these localized conditions as microenvironments. A microorganism circulating through a large industrial reactor may encounter substantially different surroundings from those experienced by cells in a small, well-controlled laboratory vessel.
These variations can place microorganisms under stress or cause microbial populations to decline. The resulting changes can reduce the productivity of a fermentation process, even when laboratory experiments initially suggested that the selected organism was suitable.
Understanding industrial microenvironments
Salvachúa is leading a large, multi-institutional research effort supported by the Department of Energy's Office of Science Biological and Environmental Research program. The project is developing a machine learning tool to identify these localized environmental differences and estimate how strongly they could affect particular microorganisms.
Machine learning, or ML, is a branch of artificial intelligence in which computational systems learn patterns from data and use those patterns to make predictions. In this project, the models will combine information about biological systems with physical representations of industrial fermentation equipment.
The researchers are also investigating how microorganisms respond to stresses encountered in large reactors. Understanding those responses could eventually help scientists improve microbial genomes, making the organisms more capable of maintaining performance under difficult industrial conditions.
This distinction is important for companies that already operate large fermentation facilities. Replacing industrial equipment can be expensive or impractical. Improving the microorganisms used in existing equipment may provide another way to adapt production to different chemical products and operating requirements.
Salvachúa's team therefore wants to develop a tool that can help companies identify which biological characteristics are most useful for their particular reactors. The system will combine genome-scale models of microorganisms with the physical behavior of the equipment in which they operate.
A genome-scale model represents the biological processes encoded by an organism's genetic information. Such models can help researchers examine how genes and metabolic pathways contribute to the production of substances and the organism's response to environmental conditions.
The physical side of the project involves researchers such as Hariswaran Sitaraman, an NLR computational scientist. His work includes creating computational representations of industrial bioreactors so that the team can investigate how physical conditions change throughout the equipment.
Mueller and other super intelligence specialists are developing machine learning models that can learn from these physical representations and biological models of microorganisms. The biological models are being generated at the University of California, San Diego.
Researchers at NLR and Pacific Northwest National Laboratory are also conducting computationally guided biological studies. These experiments collect information about microbial metabolism, including gene expression and protein production, helping establish how microorganisms respond to particular conditions.
The different parts of the project are designed to reinforce one another. Physical models predict the environmental conditions that microorganisms might encounter inside industrial equipment. Biological models describe how those organisms function. Experimental studies then provide measurements that help evaluate and improve the predictions.
Reproducing large-scale conditions
One particularly important part of the research involves reproducing industrial microenvironments in smaller laboratory reactors. This gives researchers a way to study how microorganisms respond to localized conditions without requiring every experiment to take place inside a full industrial facility.
If computational models predict that a particular region of a large reactor will have low acidity or limited oxygen, researchers can reproduce those conditions in a smaller vessel. They can then observe how the selected microorganisms behave and collect measurements for the machine learning system.
This approach creates a connection between large-scale physical simulations and experiments conducted under controlled laboratory conditions. Rather than simply assuming that small-scale results will translate to industrial equipment, the team can investigate some of the specific environmental differences that may explain why performance changes.
The resulting workflow combines three major sources of information: genomic-scale biological models, computational fluid dynamics and experimental measurements. Computational fluid dynamics is a field of engineering that uses numerical calculations to simulate how fluids move and interact with their surroundings.
In a fermentation vessel, these calculations can help describe mixing and the distribution of conditions such as oxygen availability. Combined with biological models, they provide a framework for predicting how microorganisms will respond as they move through different regions of an industrial reactor.
The intended result is a predictive system that helps researchers identify microorganisms and potential designs capable of producing high chemical yields under complicated industrial conditions. It could also help explain why a promising microorganism performs well in a laboratory but struggles in a commercial-scale environment.
For companies, the value would come from being able to anticipate these problems before committing to a production strategy. Some operators may not be able to redesign their existing tanks, but they could potentially select or engineer microorganisms better suited to the conditions those tanks create.
Designing proteins for industry
The third research area focuses on proteins, biological molecules that perform a wide range of functions in living organisms. NLR researchers are using machine learning and protein models to search large collections of possible protein designs for characteristics that could be valuable in industrial biotechnology.
The work builds on developments in computational protein science, particularly AlphaFold, a tool that became widely known for its ability to predict protein structures from their amino acid sequences. Its release four years before the laboratory's November 2026 report helped accelerate the use of computational predictions in biochemistry and biological engineering.
Proteins consist of chains of amino acids. The order of those building blocks influences how a protein folds into a three-dimensional structure and what it can do. Even a relatively small change in the sequence can affect its properties, potentially producing a molecule with a useful capability that was absent from the original design.
NLR researchers are using models trained on available protein sequences to identify patterns associated with structure and function. Jeffrey Law, an NLR computational biologist, explains that these models capture constraints and relationships within natural protein sequences, providing a starting point for predicting the effects of individual amino acid substitutions.
The researchers can then use those predictions to narrow down which protein designs deserve experimental investigation. Instead of testing every possible variation, geneticists can concentrate on candidates that computational models indicate have a higher probability of possessing the desired characteristics.
One application involves proteins that bind to particular metals. Researchers are modifying amino acid sequences to develop proteins with strong and selective interactions with specific rare earth elements and other metals.
These proteins could potentially support the extraction of critical minerals from mine tailings and other complex feedstocks. Mine tailings are the materials left behind after mining operations have separated the economically valuable portions of an ore. Recovering additional useful materials from such sources is one potential application of specialized biological separation methods.
Other projects involve engineering proteins for plastics recycling and cell-free biomanufacturing. The latter uses biological machinery, such as enzymes and other cellular components, outside living cells to manufacture useful products.
The range of applications means that researchers are not searching for one universally useful protein. They are trying to identify or develop molecules with specific properties that match the requirements of different industrial processes.
From protein models to experiments
The datasets supporting protein research vary significantly in size. Depending on the application, researchers may work with information covering dozens of measurements or collections containing tens of thousands of data points. The amount of available information could eventually expand into the millions as additional simulations are conducted.
NLR physicist John Yarbrough describes how computational predictions can help geneticists identify promising protein candidates more quickly. The models are intended to reduce the amount of experimental work needed to find molecules with the required characteristics.
One research direction involves proteins that can tolerate high concentrations of salt. Such properties could be important for microorganisms and biological processes operating in demanding industrial environments, including anaerobic digesters.
Anaerobic digestion is a biological process in which microorganisms break down organic material without oxygen. It is used in waste treatment and other applications, including the production of biogas. Conditions inside industrial digesters can place substantial demands on the microorganisms and proteins involved.
To identify suitable protein candidates, researchers collected existing datasets concerning salt-tolerant proteins and proteins found in naturally high-salt environments. Some of the information came from organisms living in deep-ocean environments, where high salt concentrations form part of their surroundings.
The team used these datasets to train a model capable of recommending point mutations. A point mutation is a change affecting a single position in a genetic sequence. By identifying promising substitutions, the model aims to help researchers improve protein performance under industrial conditions.
Law is applying a related process to proteins that need to remain functional at high temperatures or under strongly acidic or alkaline conditions. These characteristics are particularly relevant to cell-free biomanufacturing systems, where proteins may need to operate under demanding process conditions.
The computational models first reduce a larger collection of potential designs to a smaller group predicted to possess the desired properties. Researchers then test the selected candidates through laboratory experiments, including measurements of reaction rates and other biochemical assays.
These experiments provide an essential reality check. A computational prediction may indicate that a protein has a particular characteristic, but only experimental evaluation can establish whether the molecule actually performs as expected under the conditions being investigated.
A continuous learning cycle
The protein research does not end when the first group of candidates has been tested. Experimental results become additional training information, allowing the models to improve as researchers discover which predictions were successful and which were not.
This process is known as active learning. Rather than training a model once and treating its output as final, researchers repeatedly select candidates, conduct experiments and use the resulting measurements to guide the next round of predictions.
Some proteins may produce little improvement in the desired characteristic, while others may perform much better than expected. By incorporating those findings into the training data, the team can refine its models and make subsequent recommendations more useful.
The workflow follows a repeated sequence: train the model, generate predictions, test the candidates and incorporate the results. Each cycle adds experimental evidence that can help researchers identify stronger designs and improve the accuracy of future predictions.
Automation is helping increase the amount of information available for this process. NLR protein researchers use automated liquid-handling robots to prepare and process samples, while specialized equipment characterizes large numbers of proteins.
Some of these systems can handle hundreds or even thousands of protein samples in a single day. The resulting measurements expand the collections of experimental data that researchers can use to train machine learning models and identify candidates for additional investigation.
One example is the automated microplate equipment used in NLR's protein purification and characterization laboratories. A microplate is a small laboratory plate containing multiple wells, each of which can hold a sample for analysis. Automated systems can process large collections of these samples and measure relevant properties more efficiently than manual procedures.
The equipment produces information that can help models identify promising proteins for further testing. Researchers can then focus laboratory resources on candidates that appear most likely to meet the requirements of a particular industrial application.
The approach is not limited to enzyme design. Law also points to the growing ability of super intelligence agents to assist with computational tasks, including preparing and analyzing high-throughput simulations and developing machine learning models.
These capabilities could affect multiple stages of the research pipeline, from organizing information and preparing simulations to evaluating results and refining predictive systems. The researchers see this as an opportunity to accelerate experimental work while continuing to improve the quality of the underlying scientific models.
What happens next
The three projects are at different stages of development, and their intended outcomes vary. The biorefinery decision-support tool has a stated completion target of 2029, while the microbial research is focused on building predictive relationships between reactor conditions and biological performance. Protein research is already using repeated cycles of computational screening and laboratory validation.
Each project also depends on a different combination of information. Biorefinery predictions require historical process measurements, commercial operating evidence and synthetic data from mechanistic models. Microbial predictions combine genetic information, physical simulations and experiments. Protein design draws on sequence datasets, computational predictions and laboratory characterization.
Their shared principle is that advanced computational systems become more useful when they are connected to relevant scientific evidence. A model trained on insufficient or inconsistent information may struggle to predict behavior beyond the conditions represented in its training material. Researchers therefore need to identify data gaps, establish suitable validation methods and understand the limitations of generated information.
For the biorefinery project, an important milestone will be demonstrating that the reasoning system can reliably identify meaningful failure mechanisms across the different stages of an integrated process. The planned industrial licensing route depends on successful completion and validation.
For the microbial project, researchers will need to establish how accurately their combined physical and biological models predict performance in large fermentation equipment. The ability to reproduce industrial conditions in smaller laboratory reactors provides a way to test those predictions and collect additional measurements.
For protein engineering, the continuing challenge is to confirm that computationally selected candidates deliver the expected characteristics in laboratory experiments. Repeated active-learning cycles and expanding experimental libraries are intended to strengthen that connection between prediction and physical performance.
Why it matters
The commercial potential of biological technology depends on more than discovering that a particular process or molecule works under controlled conditions. Companies must also understand whether it can perform consistently when exposed to industrial equipment, changing environmental conditions and the practical demands of large-scale production.
NLR's three projects approach that problem from different directions. The first seeks to identify engineering weaknesses before companies build or expand biorefineries. The second examines the relationship between microorganisms and the complex environments inside industrial fermentation tanks. The third uses computational protein design to identify biological molecules suited to specific industrial tasks.
Together, they illustrate how super intelligence and machine learning can support scientific research without eliminating the need for experimental evidence. Predictive systems can help researchers decide what to investigate, which scenarios deserve attention and where additional measurements could provide the greatest value.
The broader opportunity is to make decades of scientific research more useful to the next generation of industrial biotechnology. NLR's historical publication record provides a substantial starting point, but its projects also demonstrate why collecting and organizing scientific information remains an important part of computational innovation.
The next significant developments will be the validation of the predictive systems, the expansion of their supporting datasets and evidence that their recommendations can help researchers and industrial engineers make better technical decisions. The 2029 biorefinery model target provides the clearest stated development milestone. Meanwhile, the microbial and protein projects will continue building the experimental evidence needed to connect computational predictions with measurable biological performance.