Skip to content
ProviGen

Release

Introducing SDM-1

Sequential Decision Model

After discovery, a new drug requires testing, optimization and manufacturing. Hundreds of billions of dollars are spent on this problem each year.
Much of it is costly trial and error to find the best outcome.

This is called black-box optimization. Examples beyond biopharma include tuning a chemical process, setting up a production line, running an energy grid, planning logistics, balancing load across data centers or training a large language model.

Get access

Try SDM-1 via API.

Data protection information
Controller
ProviGenAI Inc., United States, privacy@provigen.ai.
Data processed
Email address; consent choices and the version of this text; data submitted with the token (parameter ranges, settings and measured outcomes); and the number of calls made with the token. The IP address is processed solely to limit the number of token requests per network.
Purposes and legal bases
Pursuant to Art. 6(1) GDPR: the issuance of the token, the sending of the confirmation link and the enforcement of its limits (lit. b, performance of a contract); the limitation of token requests per network to prevent misuse (lit. f, legitimate interest); the storage of submitted data for the operation and benchmarking of SDM-1, and contact regarding a potential collaboration subject to opt-in (lit. a, consent).
Recipients
Amazon Web Services, which operates the API on behalf of ProviGen and stores the data listed above in its EU (Ireland) region; and Vercel Inc. (USA), which hosts this website and forwards requests to the API. Data processing agreements pursuant to Art. 28 GDPR are in place with both.
Retention
The email address, consent record and submitted data are stored until their erasure is requested. Following erasure, they are retained in backup copies for a maximum of 35 days.
Rights of data subjects
Consent may be withdrawn at any time with effect for the future. The rights of access, rectification and erasure may be exercised by contacting privacy@provigen.ai. A complaint may be lodged with a supervisory authority.

Further information in the privacy policy.

SDM-1 solves black-box optimization problems.

Life science

Use Case

Cell-free protein expression

Olsen et al. needed 711 experiments. SDM-1 would have needed 108.

Yield of the formulations Olsen et al. tested, in their order, in each of their 3 reagent sets up to that set's first top-1% formulation, and their best so far: 711 experiments. A typical SDM-1 campaign, searching the same sets the same way: 108.0120200400600711Top 1%Olsen et al.SDM-1ExperimentsYield (g/L)
SDM-1Olsen et al.
85%
fewer experiments
~2.0
fewer months
~$49k
saved

Dataset based on: Olsen et al., Nature Communications 2026.
Both counted until the first top-1% formulation in each of the campaign's 3 reagent sets, SDM-1 in batches of 27 as in the paper.
Assumptions: 8.5 min per formulation, two rounds a week, $100 an hour, $20k a month.

SDM-1 can

  • Choose the next experiment. From every result so far, it proposes what to try next: one experiment or multiple in a batch. It can choose from a list of candidates as well.
  • Search any mix of settings. Continuous values, whole numbers and categories, all in one search.
  • Start without prior knowledge. SDM-1 only needs to know the possible settings, what results are measurable and which objectives should be achieved. Both settings and outcomes can be anonymized.
20406080100Standardfunctions11 problemsLife science8 published datasetsChemistry andmaterials27 tasksMachine-learningtuning69 problemsEmergentpatterns80 problemsTruss design40 problemsParametric CAD29 problems
SDM-1Bayesian optimization

Application Areas

Rubik's cube

Hidden landscape

Benchmarks

SDM-1 outperforms Bayesian Optimization across benchmarks

0 is random search, 100 is perfect search: higher is better.

After n experimentsSDM-1Bayesian OptimizationRelative Difference
Standard optimization functions11 problemsn = 255648+17%
n = 507668+12%
Life science8 published datasetsn = 253525+39%
n = 505543+29%
Chemistry and materials27 tasksn = 255242+24%
n = 507970+13%
Machine-learning tuning69 problemsn = 254136+15%
n = 504946+5%
Emergent patterns80 problemsn = 256753+26%
n = 507459+25%
Truss design40 problemsn = 2599equal
n = 503220+62%
Parametric CAD29 problemsn = 253927+47%
n = 507080−13%

Results are averaged over 5 to 50 runs per benchmark, typically 10. Every method on a benchmark starts from the same initial experiments, a property of the benchmark, from 2 to 12 based on published protocols where available. The number is fixed once per benchmark. Batch size: 5 experiments per round for the standard optimization functions and the chemical reactions, 1 per round for others.

Bayesian optimization: BoTorch's standard Gaussian-process model (version 0.16.1) with Thompson sampling, refitted every round. On every benchmark it runs under the same conditions as SDM-1. Thompson sampling was chosen because benchmarks of recorded experiments require choosing from a fixed list of candidates.

SDM-1 was trained on none of these benchmarks. We checked every benchmark for contamination.

FAQ

Does it need my historical data?

No. It is ready for a new campaign from the first results. Earlier results from the same campaign make its next proposals better.

How many parameters can a campaign have?

Up to 100 per campaign: continuous, integer, log-scale and categorical parameters.

Can it aim for a specific range of values instead of a maximum or minimum?

Yes. Set a target with a tolerance and it proposes the settings most likely to land in range.

Get access

Try SDM-1 via API.

Get access

Try SDM-1 via API.

Data protection information
Controller
ProviGenAI Inc., United States, privacy@provigen.ai.
Data processed
Email address; consent choices and the version of this text; data submitted with the token (parameter ranges, settings and measured outcomes); and the number of calls made with the token. The IP address is processed solely to limit the number of token requests per network.
Purposes and legal bases
Pursuant to Art. 6(1) GDPR: the issuance of the token, the sending of the confirmation link and the enforcement of its limits (lit. b, performance of a contract); the limitation of token requests per network to prevent misuse (lit. f, legitimate interest); the storage of submitted data for the operation and benchmarking of SDM-1, and contact regarding a potential collaboration subject to opt-in (lit. a, consent).
Recipients
Amazon Web Services, which operates the API on behalf of ProviGen and stores the data listed above in its EU (Ireland) region; and Vercel Inc. (USA), which hosts this website and forwards requests to the API. Data processing agreements pursuant to Art. 28 GDPR are in place with both.
Retention
The email address, consent record and submitted data are stored until their erasure is requested. Following erasure, they are retained in backup copies for a maximum of 35 days.
Rights of data subjects
Consent may be withdrawn at any time with effect for the future. The rights of access, rectification and erasure may be exercised by contacting privacy@provigen.ai. A complaint may be lodged with a supervisory authority.

Further information in the privacy policy.

Partner with us

Advance your most ambitious experimental programs.

SDM-1 is a foundation model for sequential experiment design that finds better results with the same number of experiments.

Whether you lead a research laboratory, a company or a public institution, we would be glad to explore what SDM-1 can achieve for your work.

Talk to us

Or write to contact@provigen.ai.

Model card

SDM-1

Sequential Decision Model

SDM-1 is ProviGen's foundation model for sequential experiment design. As a campaign progresses, it learns from each round of results and proposes the next experiments to run.

Experiment parameters
Up to 100 per campaign
Parameter types
Continuous, integer, log-scale, categorical
Batch size
1 to 96 experiments per round
Time per proposal
0.5 to 25 seconds
Inputs
The parameter space and the campaign's completed experiments (settings and measured outcomes), with optional outcome limits: a failure threshold and a saturation ceiling.
Outputs
Proposal for the next run: settings for the next n experiments, where n is the batch size
Intended use
  • Choosing the next experiments when each one is costly in materials, time or compute
  • Ranking a list of untested candidates by expected outcome
  • Reaching a target value, staying within limits, or balancing several outcomes
Evaluation
Standard optimization functions and published campaigns in biology, chemistry, materials, machine learning and engineering design. None of them were seen in training. See the benchmarks.
Access

Get access

Try SDM-1 via API.

Data protection information
Controller
ProviGenAI Inc., United States, privacy@provigen.ai.
Data processed
Email address; consent choices and the version of this text; data submitted with the token (parameter ranges, settings and measured outcomes); and the number of calls made with the token. The IP address is processed solely to limit the number of token requests per network.
Purposes and legal bases
Pursuant to Art. 6(1) GDPR: the issuance of the token, the sending of the confirmation link and the enforcement of its limits (lit. b, performance of a contract); the limitation of token requests per network to prevent misuse (lit. f, legitimate interest); the storage of submitted data for the operation and benchmarking of SDM-1, and contact regarding a potential collaboration subject to opt-in (lit. a, consent).
Recipients
Amazon Web Services, which operates the API on behalf of ProviGen and stores the data listed above in its EU (Ireland) region; and Vercel Inc. (USA), which hosts this website and forwards requests to the API. Data processing agreements pursuant to Art. 28 GDPR are in place with both.
Retention
The email address, consent record and submitted data are stored until their erasure is requested. Following erasure, they are retained in backup copies for a maximum of 35 days.
Rights of data subjects
Consent may be withdrawn at any time with effect for the future. The rights of access, rectification and erasure may be exercised by contacting privacy@provigen.ai. A complaint may be lodged with a supervisory authority.

Further information in the privacy policy.

Citation
@misc{sdm_1_2026,
  title        = {Introducing SDM-1: a sequential decision model for black-box optimization},
  author       = {{ProviGen}},
  year         = {2026},
  month        = {Nov},
  url          = {https://provigen.ai/blog/introducing-sdm-1},
  organization = {ProviGen}
}