☰

Dunia Innovation

From screen to bench: closing the synthesis gap

Industry
Deep Tech / AI for Materials Discovery
Headquarters
Berlin, Germany
Public information as of
January 2026

A4BEE prepared this analysis from publicly available sources. It reflects our own reading of Dunia Innovation's published strategy and is not endorsed by, or produced in cooperation with, Dunia Innovation. Company website

Strategic priorities

Dunia Innovation, headquartered in Berlin and founded in 2022, is building self-driving laboratories that combine AI-designed materials with autonomous robotic synthesis and testing. The company raised $11.5 million in October 2024 in a round co-led by Elaia and Redalpine, with participation from the European Innovation Council and others, to scale toward its first GigaLab facility by 2026, designed to design, synthesize and test up to 1,000 materials per day.

The strategic frame is accelerating materials discovery from a roughly 20-year cycle to under three years in electroactive materials and electrocatalysts for water electrolysis and CO2 reduction. Public reporting cites a 1-million-validated-materials-data-point target for 2026, alignment with Saudi Vision 2030 following Dunia's selection as a winner of the FII9 Innovators Pitch 2025 in Riyadh, and stated applicability to green hydrogen, ammonia and CO2-to-chemicals value chains.

The practical constraint sits between the AI design layer and the physical lab. The CEO has publicly named the synthesis bottleneck — software can propose 80,000 candidate materials in weeks while physical making and testing remains limited by traditional lab infrastructure — and the wider field has identified that AI models trained on incomplete or inaccurate data inherit that gap. Translating every instrument reading into structured, traceable data is what makes autonomous material design credible.

Operating priorities line up accordingly: a GigaLab build by 2026, modular equipment architectures compatible with the MTP standard, computer-vision quality control, and a planned expansion of the Cloud Lab model across the MENA region, each of which depends on lab data being captured as data rather than transcribed from screens.

Challenges we see

  • Digital R&D Operations

    Producing traceable data from every synthesis step

    Dunia's CEO has publicly described the synthesis bottleneck: AI can propose tens of thousands of material candidates in weeks, while the physical making and testing still depend on traditional laboratory infrastructure. Public reporting also notes that many AI models are trained on incomplete or inaccurate data, which feeds back into the discovery loop.

    Where the materials record is split between paper notes, isolated spreadsheets and instrument-local storage, the same experiment is hard to repeat, harder to attribute, and difficult to feed into a model that others can reuse.

  • Operations Manufacturing

    Stabilising electrochemical CO2 reduction at industrial current densities

    Dunia's focus on CO2 electroreduction to multicarbon products covers well-documented process engineering challenges in the literature, including membrane flooding from salt precipitate accumulation, sluggish ion transport through charge-selective membranes, and catalyst degradation. The literature notes that gas diffusion electrode systems for multicarbon products commonly fail after around thirty hours of operation, with energy efficiency for CO reduction to those products still below 40 percent.

    When stability constrains the run length that can be tested in a day, the bottleneck shifts from how many candidates can be made to how long each candidate can stay on test, which changes what the autonomous loop needs to measure.

  • Digital Integration

    Connecting shop-floor instruments to the management stack

    Public sources on the IT/OT architecture gap describe a bifurcation between high-level management systems such as SAP or INTENSE and the actual shop floor equipment, where humans act as manual middleware, typing values from machine displays into digital ledgers. Many R&D centres mix modern lines with 1990s-era equipment that does not expose real-time data.

    Where instruments and management systems sit in separate estates, every reading is a separate transfer with its own chance of being lost, delayed, or mistyped before it reaches the records that the rest of the business reads from.

  • Operations Organizational

    Building the specialist skill base the autonomous loop depends on

    Public industry reporting indicates that 57 percent of respondents identified workforce skill gaps as the largest impediment to successful automation, with heavy reliance on manual tasks such as pipetting, agar media preparation and visual colony counting in research environments where budgets limit the adoption of top-tier robotics.

    When the team profile is built around manual techniques, the natural next step is automation delivered in forms the team can run, so the skills to operate and maintain the system grow inside the organisation rather than being imported one project at a time.

  • Compliance Regulatory

    Documenting scale-up evidence for regulatory bodies

    Moving from one-litre lab-scale batches to 500-litre and larger reactors introduces biological and chemical unpredictability. Regulatory routes such as EFSA Novel Food approvals and medical device certifications require traceable data across the development record, and decisions documented across multiple sites often need to be assembled from many different systems.

    When the evidence base for a regulatory submission lives across instruments, notebooks and spreadsheets, the time between an answer being true and the answer being provable becomes a property of the documentation pipeline rather than the experiment.

Opportunities, by urgency and business impact

Each bubble is one opportunity, numbered to match the list below. Further right means it bites sooner; higher means a bigger effect on the business. A bigger bubble means a bigger implementation effort.

Source: A4BEE analysis of public sources

  1. Capturing lab instrument data as a continuous record

    Critical research data at materials discovery labs is still commonly recorded on paper or in isolated spreadsheets, which blocks AI models from training on accurate, complete datasets and prevents root cause analysis during scale-up.

    A unified data gateway that streams readings from synthesis robots, characterisation instruments and electrochemical cells into a single dataset — tagged with sample, method and instrument identity — gives the discovery model a continuous evidence base and gives engineers a record they can replay.

    • Dunia.ai — Platform and Pipeline pages (dunia.ai/platform, dunia.ai/pipeline)
    • Public interview with Dunia's CEO, Silicon Canals, 2024

  2. Bridging the IT/OT architecture gap across modern and legacy equipment

    R&D centres typically run a mix of modern lines and 1990s-era equipment. Data is held inside instruments or vendor systems, and operators transcribe values into management systems by hand.

    Non-invasive signal acquisition from existing PLCs and instruments, combined with an ontology-based data model that normalises different equipment generations, lets shop-floor data reach the management stack as data rather than retyping.

    • A4BEE service reference: Industrial Data Platform
    • Public sources on IT/OT architecture gaps in R&D environments

  3. Building MTP-compliant modules for the 2026 GigaLab

    Rigid infrastructure with proprietary vendor lock-in prevents rapid adaptation to new material classes. Long downtimes are needed for facility upgrades, and modular research goals like a 1,000-materials-per-day GigaLab by 2026 require equipment that can be swapped and expanded without rewriting control logic.

    Adopting the MTP (Module Type Package) standard from VDI/VDE/NAMUR 2658 — with OPC UA servers and state-based interfaces on each Process Equipment Assembly — turns equipment swap-outs into plug-and-produce operations and keeps recipe management consistent as new modules arrive.

    • Dunia 2024 funding announcement and Elaia portfolio entry
    • A4BEE article: Accelerating Lab and Manufacturing Operations with MTP

  4. Automating visual inspection and QC through ML vision

    Manual pipetting, visual colony counting and human observation introduce variability and limit throughput in quality control, with no permanent digital record of the visual judgements made during an experiment.

    Camera-based ML algorithms — object detection on microscope images of catalyst surfaces and bioreactor imagery, plus foam detection in reactors — produce a permanent digital record, run continuously, and replace routine human observation with a system the team can query later.

    • Dunia 2024 funding announcement
    • A4BEE case: Microalgae Cells Detection and State Classification ML Model Based on Microscope Images

  5. Protecting material IP and enabling Cloud Lab collaboration across regions

    High-value material IP must be protected across geographically dispersed facilities and during collaboration with external partners, especially as Dunia plans Cloud Lab expansion across the MENA region following the FII9 Innovators Pitch 2025 selection in Riyadh.

    Zero-trust identity-based access, end-to-end encryption of research data, and segmented remote-access paths let researchers, partners and remote support engineers work with the same instruments without material formulations leaving their intended audience.

    • Entrepreneur Middle East, Inside FII9: Dunia Innovations, 2025
    • A4BEE article: Zero Trust Security Principles

What we'd propose

  • Digital Lab

    Data gateway for the autonomous materials lab

    A unified data gateway that streams readings from synthesis robots, characterisation instruments and electrochemical cells into a single dataset, so that every experiment leaves a traceable record from sample to result.

    • Multi-instrument data acquisition

      Getting data off the lab instruments

      Connect synthesis robots, chromatography systems, balances and electrochemical cells through OPC UA (Open Platform Communications Unified Architecture) or MQTT so process values leave each device in a documented, vendor-neutral form rather than being read from a screen and retyped.

    • Link to LIMS and research notebooks

      One record per experiment

      Route instrument streams into the laboratory information system and electronic notebook so a sample, its method and its results share a single identifier that can be replayed at any time.

    • GxP-compliant data pipelines

      Traceability that survives an audit

      Apply FDA and GMP-aware data lineage — signed records, versioned methods and an audit trail from instrument to result — so the dataset is ready for regulatory inspection without retrospective reconstruction.

    • Every reading carries an instrument identity, a method version and a timestamp, so an experiment can be replayed rather than reconstructed.
    • Material property datasets can be fed directly into the AI design loop without manual transcription.
    • The same dataset serves R&D, scale-up engineering and regulatory reporting instead of three separate extracts.

    We've done this before

    See how we do Digital Lab
  • Enterprise AI

    IT/OT integration for a mixed-generation R&D fleet

    A reference architecture and integration layer that brings modern instruments and legacy 1990s-era equipment onto one data plane, so shop-floor readings reach the management stack as data rather than as retyped numbers.

    • Non-invasive signal acquisition

      Reading data from older controllers

      Add signal-acquisition gateways to legacy PLCs and instruments to capture process values without disturbing the controlling software, so equipment from any generation contributes to the same data stream.

    • Ontology-based data model

      One vocabulary across equipment eras

      Define equipment, batch, sample, parameter and result as explicit entities with agreed relationships, so a query written once returns comparable answers across modern and legacy instruments instead of one dialect per generation.

    • Workflow integration with management systems

      Readings that trigger the next step

      Connect the data plane to SAP-class management systems so shop-floor readings update production and quality records, and so threshold crossings schedule the next action automatically rather than waiting for a manual entry.

    • Legacy and modern equipment contribute to one data stream without retiring existing machinery.
    • Cross-vendor and cross-generation comparison becomes a query rather than a translation exercise.
    • The same management-system integration is reusable at every future site.

    We've done this before

    See how we do Enterprise AI
  • Digital CDMO

    MTP-compliant modular architecture for the GigaLab build

    Process Equipment Assemblies built to the MTP (Module Type Package) standard with OPC UA servers and state-based interfaces, integrated through a Process Orchestration Layer so the GigaLab can be expanded, reconfigured or serviced module by module.

    • MTP-compliant Process Equipment Assemblies

      Modules that share a language

      Specify new equipment as MTP-compliant Process Equipment Assemblies with OPC UA servers and state-based interfaces to VDI/VDE/NAMUR 2658, so any vendor's module can be brought into the orchestration layer without custom adaptation.

    • Process Orchestration Layer

      One control plane above the modules

      Build a Process Orchestration Layer that treats each Process Equipment Assembly as a state-driven unit, so recipe management, batch tracking and remote operations all reference a single interface rather than separate vendor controllers.

    • Vendor-agnostic implementation guidance

      A blueprint both sides can follow

      Guide internal teams and equipment vendors through the modular design choices, so module delivery, validation and integration follow a shared pattern that compounds with each new module.

    • Equipment replacement or expansion is a swap of one module, not a rewrite of the control system.
    • New material classes arrive faster because the orchestration layer does not need to be rebuilt for each one.
    • Vendor selection becomes a recurring decision rather than a one-time commitment.

    We've done this before

    See how we do Digital CDMO
  • Digital Lab

    ML vision and digital twins for electrochemical QC

    Camera-based ML models for catalyst and cell-state inspection together with digital twin simulations of electrochemical cells, so visual judgements are captured as data and parameter studies are run in software before they are run in hardware.

    • Microscope and reactor vision

      Visual judgement captured as data

      Deploy object-detection models on catalyst surface images and foam detection on reactor footage, so each visual judgement becomes a queryable record at the same identity as the underlying experiment.

    • Digital twin of the electrochemical cell

      Run the study in software first

      Build a physics-informed simulation of the electrochemical cell — current density, electrolyte composition, gas diffusion behaviour — so the parameter study that would take weeks on the bench is screened in software before the bench is engaged.

    • Adaptive control from vision inputs

      Cameras that close the loop

      Connect ML vision outputs to the equipment control layer so threshold crossings adjust process parameters without waiting for an operator to watch a screen.

    • Visual inspections produce a permanent record the team can replay, search and compare across runs.
    • Parameter studies move from weeks of bench time to hours of in-silico screening.
    • Threshold crossings adjust the process without depending on an operator watching a screen.

    We've done this before

    See how we do Digital Lab

Digital maturity: today and target

Scored out of 100 across six dimensions. The target is what Dunia Innovation's own published ambition implies — not a perfect score.

Source: A4BEE analysis of public sources
Data Lineage & Traceability 35 → 95
Public reporting on the synthesis bottleneck notes that critical research data is often recorded on paper or in isolated spreadsheets, which limits traceability from sample origin to industrial validation. The stated 2026 target is fully digital lineage.
QC Automation 25 → 90
Public industry reporting cites heavy reliance on manual tasks such as pipetting, agar media preparation and visual colony counting, with budget constraints limiting adoption of top-tier robotics. The target is camera-based ML inspection operating on a continuous record.
IT/OT Integration 30 → 85
Public sources describe a bifurcation between SAP-class management systems and shop-floor PLCs, with manual data entry bridging them. The target is unified real-time flow from hardware sensors through an ontology-based model to the management stack.
Scaling Predictability 40 → 90
Moving from 1L lab-scale to 500L and larger reactors introduces unpredictability, and process literature notes that multicarbon-product electrochemical systems fail after roughly 30 hours. The target is Quality by Design through digital twin simulation before scale-up is committed.
Facility Modularity 45 → 85
Current R&D environments are described as rigid and proprietary. The 2026 GigaLab target implies modular, vendor-agnostic equipment built to the MTP standard so new material classes can be added without a control-system rewrite.
Regulatory Readiness 35 → 90
Regulatory routes such as EFSA Novel Food approvals and medical device certifications require traceable data across the development record. The target is an audit-ready digital record generated continuously rather than assembled on demand.

Check this yourself

Our Service Portal has free self-assessments and market comparisons. These are the ones that line up with what we've read above — no sales call required.

Think we've read this right?

Talk to us

Related reading

This is an independent analysis prepared by A4BEE from publicly available information as of January 2026. It reflects A4BEE's own interpretation and opinion, is not affiliated with, endorsed by, or verified with Dunia Innovation, and may be incomplete or inaccurate. All company names and trademarks are the property of their respective owners. To request a correction or removal, contact [email protected].