01
Food & nutrition
Quality from harvest to product.
Compare growing and processing choices against nutritional composition, shelf life, sensory quality and usable output.
Intelligence for agricultural systems.
Discovering new value
in living systems.
Glassbiome is building intelligence infrastructure for agriculture and biological production. Our mission is to advance food security, improve nutrition and make better use of biological resources.
Powered by Century AG
Intelligence for
physical discovery.
Century is our AI research and decision platform. It helps identify promising opportunities in agriculture and biological production, and assess the evidence behind them.
Our work spans cultivation, ingredient processing and resource use. Physical experiments will test the predictions that emerge.
The research foundation
15 September 2026 snapshot- Research records indexed
- 238,812
- Parameters explored
- 49
- Crops & organisms explored
- 98
- Literature-scan findings
- 954
What the numbers cover
- Research records indexed
- Distinct document metadata records, deduplicated by identifiers, including papers and other research records. Indexing is separate from full-text acquisition and in-depth scientific review.
- Parameters explored
- Distinct environment and process parameter entries in completed literature searches.
- Crops & organisms explored
- Crop and organism entries in completed searches, including varieties, algae and fungi; not a count of independently tested species.
- Literature-scan findings
- Research observations catalogued in the findings portfolio, spanning different review stages. This is not a count of independently validated discoveries or proven commercial outcomes.
Research across industries
Explore the research.
Crops, processes and products.
Connected by testable questions.
Research landscape
Select an industry or research theme.
Question to investigate
Which properties make a biological ingredient useful in a product?
- Define the product requirement
- Compare material properties
- Test application performance
An ingredient comparison tied to a specific application.
The Century research engine
Discovery, measurement
and continuous learning.
- 01 / Focus
Start with a useful question.
Identify a production need, a desired property or an opportunity to improve.
- 02 / Explore
Understand the possibilities.
Examine the evidence and the questions it leaves open.
- 03 / Evaluate
Assess what matters.
Weigh potential value, practical constraints and confidence in the result.
- 04 / Test
Define the next study.
Develop focused tests for the most promising opportunities.
Engineering progress
Efficient computation. Repeatable results.
83%less CPU time
Experiment selection, with unchanged recommendations at 120 tested checkpoints.
66%smaller input
A greenhouse-model input, retaining all observations.
Consistent results across machines.
Research comparisons checked for repeatability.
- Repeated runs
- Matching scientific resultsVerified on completed processing comparisons
- Separate machines
- Numerical agreementChecked against a predefined tolerance
What was measured
October 2026 development benchmarks. Experiment-selection CPU time fell from 460.4 to 79.2 seconds across 20 paired computational runs. Recommendations matched at all 120 checkpoints; measurement order sometimes differed. This measures selection computation, not total research time or crop performance.
One greenhouse-model input shrank from 2.99 to 1.03 MB, with all observations retained and exact reconstruction verified. This measures serialized input size, not memory use or inference cost. MB labels use decimal units.
Completed processing comparisons reproduced their scientific outputs on repeat runs. A separate check found numerical agreement across two machines within a predefined tolerance. These checks establish software consistency, not independent biological confirmation.
Our focus areas
From crops to useful products.
Start with a production need.
Define the required performance.
02
Agriculture & resources
Crop performance and resource demand.
Evaluate how light, water, nutrients and climate affect crop output, composition and resource demand.
03
Beauty & personal care
Ingredient performance in formulation.
Study how cultivation, extraction and storage affect composition, stability and performance in the finished product.
04
Circular industries
Recovery, reuse and material value.
Assess byproducts and underused biomass against recovery yield, material quality, process costs and existing uses.
Prediction in practice
Finding the window
worth testing.
How long should a finishing-light treatment last?
A published red-lettuce study lets us test what different models can recover.
Leaf pigment concentrationmg CYA / g dry tissue
The four-day treatment
Both models fit the other eight settings.
- Predicted
- 12.42
- Published measurement
- 12.91
- Linear interpolation
- 8.94
36%
lower prediction error
Quadratic model vs linear interpolation.
Across seven held-out settings in this study.
Cammarisano et al., 2020 ↗Retrospective, within one study.
No independently validated optimum.
About this comparison
How concentration changes with treatment duration
Leaf pigment concentrationmg CYA / g dry tissue
- Published means
- Quadratic model
- Linear interpolation
This is a retrospective comparison using a previously inspected published red-lettuce study, not a prospectively blinded prediction. Each assessed treatment setting was omitted from fitting. The chart shows the four-day case, estimated from the other eight settings.
The 36% reduction compares root mean squared error (RMSE) across seven interior settings: 1.026 for the quadratic model versus 1.608 for linear interpolation, in mg CYA/g dry tissue. Five models were evaluated; boundary settings were not scored. This result does not establish a general prediction advantage.
CYA denotes cyanidin equivalents, a pigment assay. Points show published means and whiskers show reported standard errors (SEM), not prediction intervals. Four days corresponds to 72 lit hours under an 18-hour daily photoperiod. The duration axis uses log spacing. Treatment timing and cumulative light changed together, so duration was not isolated.
The curve's unmeasured peak is not an established growing window. The advantage did not extend to fresh-mass prediction. Pigment and mass were measured on separate plants; no recovered compound yield or production benefit is established. Independent studies and physical trials are still needed.
Research in practice
What grows.
What it contains.
Five nutrient recipes.
One published lettuce experiment.
Relative to the 0% recipe
- Shoot mass
- Leaf nitrate
Ammonium share of supplied nitrogen
75% ammonium recipe
Each outcome as a share of its 0% recipe value
- Shoot mass
- 96%
- Leaf nitrate concentration
- 36%
Much lower mean nitrate, with mean shoot mass close to the reference.
About this study
Lettuce under 24-hour daily light, harvested after four weeks. Dots compare approximate published means, with each outcome normalized to its own 0% recipe mean. Links join two separate outcomes for one recipe, not paired plant measurements, error bars or a combined benefit score. The initial selection is an inspection state, not a recommendation. Repeat testing is needed to assess the pattern.
Values were read from Figures 2A and 3B and are approximate. Values originate in one figure-digitization pass. Source and arithmetic checks do not establish independent measurement or biological confirmation. Shoot fresh mass and sampled-leaf nitrate concentration are different measurements; they do not establish whole-plant nitrate content or saleable yield. The paper reports n = 5–6, with exact counts for each treatment unresolved in this reconstruction.
These five recipes were tested within the same 24-hour-light run. Other ions and electrical conductivity also varied. The table reports approximate means ± reported standard errors, not confidence intervals. Uncertainty is not plotted on the normalized chart. This comparison does not establish an isolated ammonium effect, an optimum, or a production recommendation.
| Ammonium share | Shoot mass (g/plant) | Leaf nitrate (mg/100 g FW) |
|---|---|---|
| 0% | 157.8 ± 8.8 | 373.3 ± 16.0 |
| 25% | 158.7 ± 10.0 | 310.2 ± 7.5 |
| 50% | 159.3 ± 13.4 | 186.1 ± 6.4 |
| 75% | 151.9 ± 8.8 | 135.8 ± 7.0 |
| 100% | 94.3 ± 7.7 | 19.3 ± 3.2 |
Leadership
The co-founders.
Recent research
& engineering.
Explore the results and the capabilities behind them.
Group-averaged mean absolute error 275 held identity groups
Scientific result
A stronger molecular prediction baseline.
Adding graph features to an enriched chemical-formula model lowered average error across 284 supported records.
Retrospective development. One source; outcomes previously inspected. Gains vary by molecule.
Comparison & scope
Both models used the same supported records and fitting, calibration and test partitions. The error fell from 35.991475 to 28.268252 in the source's reported retention-index coordinate. The 284 supported test records represent 275 identity groups; one original test record remains a parser refusal.
Error worsened in 114 of 275 groups compared with enriched formula alone. The source outcomes had already been inspected during development. The gain does not identify graph connectivity as its cause, establish isomer identification, or validate transfer to another laboratory or chromatographic method.
Source: Published PTFI molecular library .
Fixed reference scalemmol/100 g
24 categories × 2 reference rules × 3 dependent allocations
Measurement engineering
Change the units.
Keep the meaning.
Two prediction pipelines preserved equivalent assay predictions when reported values and reference scales changed units together.
Same assay and sample basis. The toggle converts a declared reference scale; it is not a measured sample.
What was checked
All 144 paired comparisons passed the stated numerical tolerances after fitting at both mmol/100 g and µmol/100 g. The check covered two reference-scale rules, 24 food categories and three dependent disclosure allocations per category.
This checks the declared conversion for the same endpoint and sample basis. It does not make different assays equivalent or establish cross-laboratory calibration.
Source: Carlsen et al., 2010 .
53 reported means 1,449 source entries checked
Research engineering
A workflow that travels between studies.
Peel, herb and beer studies entered one maintained predictor, without custom model code for each study.
Conventional controls led these comparisons. Data reuse does not establish cross-study predictive transfer.
What was reused
Peel, herb and beer study packets entered one engine without custom model code for each study. The 1,449 checked entries include missing and ambiguous compound fields; they are not all measured quantities. Reported means are not established independent biological samples.
All 380 model responses returned. Total-phenol regression led the peel and herb FRAP comparisons; the ordinary median led the beer ABTS and DPPH comparisons. The composition model was not adopted as the general winner.
Research & commercial partnerships
Bring a production
decision into focus.
Work with us on reliable food production, ingredient quality or resource recovery.
Define the objective, evaluate the evidence and design the validation programme.