Synthetic Systems AI: From Generating Content to Generating Systems

A new class of artificial intelligence for understanding, inventing and evaluating complex systems

Generative artificial intelligence has demonstrated that machines can produce text, images, software code, audio and video.

This is significant, but it remains focused primarily on the generation of representations.

The next step is not simply to generate more content. It is to generate, analyse and evaluate the systems that the content describes.

We propose the term Synthetic Systems AI for this emerging capability:

Synthetic Systems AI is artificial intelligence that reconstructs, generates, simulates and evaluates alternative system architectures.

Its primary objects are not sentences, images or isolated code fragments. They are systems composed of interacting components, states, processes, constraints and feedback structures.

Synthetic Systems AI therefore represents a shift:

Generative AI generates content. Synthetic Systems AI generates systems.

Why systems require a different kind of intelligence

A complex technical system cannot be understood as a collection of independent components.

Its behaviour emerges from interactions among:

physical components,
sensors and actuators,
software and control logic,
data and communication flows,
system states,
feedback loops,
environmental conditions,
safety constraints,
economic requirements,
legal and intellectual-property conditions.

Changing a single component may alter the behaviour of the entire system. A technically superior component can make the overall system less stable. Two individually mature technologies may create new risks or dependencies when they are combined.

Conventional generative AI can describe such systems. It can produce reports, code or diagrams about them. But describing a system is not the same as representing its structure, generating technically meaningful alternatives or predicting the consequences of architectural changes.

Synthetic Systems AI treats the system itself as the object of computation.

A formal view of synthetic systems

A system may be represented abstractly as

$$
S=(V,E,X,P,C,F)
$$

where:

$V$ represents components and subsystems,
$E$ represents physical, informational and functional relationships,
$X$ represents states and state variables,
$P$ represents parameters,
$C$ represents technical, economic, legal and operational constraints,
$F$ represents functions and intended technical effects.

A Synthetic Systems AI transforms an existing system $S$ into candidate alternatives:

$$S’_1,S’_2,\ldots,S’_n =
G(S,K,O)$$

where:

$K$ is an engineering knowledge base,
$O$ contains objectives and constraints,
$G$ is a system-generation process.

The generated alternatives are then evaluated according to multiple criteria:

$$
\begin{aligned}
Q(S‘) =&+ w_1Q_{\mathrm{performance}}\\
&+ w_2Q_{\mathrm{robustness}} \\
&+ w_3Q_{\mathrm{safety}} \\
&+ w_4Q_{\mathrm{feasibility}} \\
&+ w_5Q_{\mathrm{novelty}} \\
&+ w_6Q_{\mathrm{economic\ value}}.\end{aligned}
$$

This is not merely text generation followed by a plausibility judgement. The objective is to create executable system hypotheses that can be inspected, simulated, compared and eventually implemented.

The four capabilities of Synthetic Systems AI

1. System reconstruction

Technical knowledge is currently distributed across patents, scientific papers, engineering documentation, standards, product specifications, source code and operational data.

Synthetic Systems AI reconstructs systems from these heterogeneous sources. It identifies:

components and their functions,
inputs and outputs,
system states,
causal and functional relationships,
technical constraints,
claimed effects,
failure modes,
alternative implementations.

The result is not merely a document collection. It is an engineering knowledge graph in which technical entities and relationships remain linked to their original evidence.

For example, the graph may represent that:

a patent family contains a particular independent claim,
the claim requires a state estimator,
the estimator receives sensor data,
its result is supplied to a predictive controller,
the controller restricts an actuator,
the claimed technical effect is improved system stability.

Each relation should retain provenance:

source document,
section or claim,
extraction method,
confidence level,
review status.

This provides the foundation for traceable reasoning rather than ungrounded generation.

2. System synthesis

Once systems are represented structurally, they can be modified systematically.

Synthetic Systems AI may generate alternatives by:

replacing components,
removing dependencies,
combining mechanisms from different domains,
changing control hierarchies,
introducing redundancy,
modifying feedback structures,
changing the allocation of functions between hardware and software,
introducing alternative sensing or state-estimation methods,
adding safety and fallback mechanisms.

Consider a simplified architecture:

Sensor
→ state estimation
→ digital twin
→ controller
→ actuator

Possible synthetic variants include:

Sensor fusion
→ probabilistic state estimation
→ ensemble of system models
→ robust predictive controller
→ actuator

or:

Virtual sensor
→ hybrid physical and data-driven model
→ runtime safety filter
→ learning controller
→ actuator

The important output is not the wording of these alternatives. It is the explicit change in architecture and the consequences that follow from it.

3. Simulation and validation

A generated system is initially a hypothesis.

An LLM may explain why an architecture appears plausible, but plausibility alone is insufficient for engineering. Synthetic Systems AI must therefore connect generative reasoning to executable models.

Depending on the problem, validation may use:

differential-equation models,
discrete-event simulation,
agent-based models,
Monte Carlo simulation,
model predictive control,
surrogate models,
formal verification,
hardware-in-the-loop testing,
digital twins.

This creates a sequence:

Generated architecture
→ formal system model
→ simulated scenarios
→ performance and risk metrics
→ revised architecture

The resulting information can be classified as:

hypothesised, when produced by generative reasoning,
derived, when supported by formal analysis,
simulated, when demonstrated in an executable model,
experimentally validated, when confirmed using physical or operational data.

This distinction is essential. Synthetic Systems AI should not conceal uncertainty behind fluent technical language.

4. System selection and evolution

The technically best system is not always the economically relevant system.

System selection may also depend on:

development cost,
time to market,
component availability,
interoperability,
regulatory requirements,
patent exposure,
licensing conditions,
difficulty of alternative implementations,
strategic relevance of particular markets.

Synthetic Systems AI therefore evaluates not only whether a system can work, but whether it creates a viable technological and economic position.

Over time, the system can learn from:

simulation results,
rejected alternatives,
patent examination histories,
field failures,
product adoption,
licensing outcomes,
expert corrections.

The engineering knowledge graph becomes both a memory of existing technology and a basis for generating its successors.

Patents as structured engineering knowledge

Patents are a particularly valuable input for Synthetic Systems AI.

A patent document contains more than a description of an invention. Depending on the jurisdiction and procedural stage, it may provide:

technical background,
system components,
alternative embodiments,
claimed combinations of features,
cited prior art,
examiner objections,
amendments to claims,
legal status,
relationships to national and international patent-family members.

Patent classifications, citation networks and family structures already form a partial graph. However, this graph describes primarily the organisation of patent documents.

Synthetic Systems AI adds an engineering layer:

Patent family
→ addresses technical problem
→ requires component
→ processes state variable
→ applies control mechanism
→ produces technical effect
→ depends on interface
→ competes with alternative architecture

The most valuable information is often contained in the independent claims and in the evolution of those claims during examination.

A system can compare:

the originally claimed invention,
prior art cited against it,
features rejected as known or obvious,
features added during prosecution,
the final granted claim.

This creates a claim-evolution graph that shows which technical distinction survived examination.

From patent landscapes to future system landscapes

Traditional patent analytics asks:

Who owns patents in a field?
How many applications are being filed?
Which patents cite one another?
Which companies are increasing their activity?

Synthetic Systems AI can ask a different class of questions:

Which currently separate technologies are likely to become interdependent?
Which interfaces will be required when they are combined?
Which technical dependencies are not yet visible in existing products?
Which existing patents may become strategically relevant in a future architecture?
Which combinations have not yet been protected?
Which integration mechanisms could become de facto standards?

Suppose two patent families describe:

Family A:
state estimation using digital twins

Family B:
runtime safety filters for autonomous control

The future strategic opportunity may not lie in either family individually. It may lie in the mechanism that connects them:

State estimate
→ predicted reachable-state set
→ dynamically admissible action set
→ runtime safety filter

This connecting mechanism may become a technological coordination point.

If it is technically sound, legally stable and available under reasonable licensing conditions, companies may adopt it rather than spend years developing and defending an alternative.

The objective is therefore not necessarily to create an unavoidable blocking patent. A more powerful strategy may be to create a licensed coordination point:

A sufficiently good, early, well-documented and legally predictable implementation that is economically preferable to designing around it.

Prospective invention engineering

Synthetic Systems AI can extend beyond currently visible white spaces.

It can analyse:

scientific development paths,
hardware performance trends,
cost trajectories,
regulatory changes,
emerging system requirements,
dependencies between technology clusters.

From these it can construct plausible future systems and work backwards to identify missing technologies.

This process may be described as prospective invention engineering.

The system begins with a plausible future condition:

Industrial plants will operate with substantially greater autonomy.

It then asks:

Which technical capabilities must exist?
Which new failure modes will emerge?
Which interfaces will become critical?
Which safety mechanisms will be required?
Which of these mechanisms can already be specified and simulated?
Which technically executable solutions remain unprotected?

The objective is not to predict a single future. It is to model several technically consistent development corridors and identify recurring structural dependencies.

A patentable invention must still be sufficiently disclosed, technically workable and non-obvious. Synthetic Systems AI cannot convert a vague future ambition into a valid patent merely by generating elaborate language.

It can, however, help transform a future requirement into:

a concrete architecture,
explicit components,
data and control flows,
mathematical procedures,
testable effects,
alternative embodiments,
simulation evidence.

Latent strategic patents

Synthetic Systems AI can also identify value in patents that already exist.

A patent may have been filed for a narrow or early application and remain commercially insignificant for years. Its underlying mechanism may later become relevant when previously separate technology domains converge.

Such patents may be called latent strategic patents.

For example, a patent originally directed at heterogeneous scientific computing may contain a general mechanism for:

distributing computational graphs across different accelerator types,
synchronising separate memory domains,
managing cross-device uncertainty,
routing work according to hardware-specific constraints.

Its future relevance may arise when new combinations of accelerators become commercially important.

The value does not come from keyword similarity. It comes from recognising a structural equivalence between:

the originally disclosed mechanism,
the requirements of a later system architecture.

This is a natural task for a combination of knowledge graphs, graph-based retrieval, semantic search and system simulation.

Synthetic Engineering Data

The same infrastructure can produce a further asset: Synthetic Engineering Data.

A system architecture can be changed in a controlled way:

Original:
single state model

Mutation:
weighted ensemble of state models

The expected effects can then be represented structurally:

Potential advantages:
- greater robustness to model misspecification
- explicit representation of model uncertainty

Potential disadvantages:
- higher computational cost
- more complex model calibration

Required validation:
- estimation error under regime changes
- runtime latency
- behaviour under conflicting model outputs

Where possible, these effects can be simulated and converted into measured labels.

This produces training data for models that must learn:

technical cause-and-effect relationships,
system trade-offs,
claim-feature extraction,
architecture comparison,
patent retrieval,
design-around generation,
engineering question answering.

The strongest product is not a collection of automatically generated patent-like texts.

It is a dataset combining:

real technical evidence,
structured system representations,
controlled counterfactual modifications,
simulated outcomes,
uncertainty annotations,
provenance.

Such data may be used to train engineering copilots, patent-analysis systems, specialised retrieval models and future system-synthesis models.

Synthetic Systems AI is not simply Generative Systems Engineering

There are already adjacent concepts.

Generative Systems Engineering applies generative AI to systems-engineering processes such as requirements analysis, model creation, documentation and lifecycle support.

Synthetic System Intelligence has also been discussed as intelligence and knowledge generated within or by cyber-physical systems and potentially transferred as an industrial asset.

Synthetic Systems AI differs in its central object and scope:

It focuses on the reconstruction, generation, simulation and strategic evaluation of alternative systems.

It integrates:

engineering knowledge graphs,
generative models,
system simulation,
patent intelligence,
prospective technology analysis,
synthetic engineering data.

The aim is not merely to accelerate an existing engineering workflow. It is to build an intelligence infrastructure that can explore the space of technically possible systems.

A different conception of intelligence

Synthetic Systems AI also implies a broader view of intelligence.

Intelligence is often treated as a scalar:

narrow intelligence
→ human-level intelligence
→ superintelligence

This assumes that intelligence is a single capability that can be increased linearly.

A more useful interpretation is relational:

$$
I=I(A,E,G,T)
$$

where intelligence depends on:

agent $A$,
environment $E$,
goals $G$,
relevant time scale $T$.

Human intelligence is strongly adapted to a specific physical, biological and social environment. Language, narrative reasoning and social inference are highly valuable in that environment.

A different intelligence may operate primarily on:

system graphs,
state spaces,
probability distributions,
reachable sets,
feedback structures,
alternative futures,
technical constraints.

It would not necessarily be a more advanced version of human intelligence. It would be a different form of intelligence with different elementary objects and operations.

Synthetic Systems AI should therefore not be judged only by whether it communicates like a human. Its defining capability is whether it can produce useful transformations in the space of systems.

The RCS perspective

Renner Complex Systems develops methods for modelling, simulation and artificial intelligence in complex technical environments.

For RCS, Synthetic Systems AI is not intended as a speculative label detached from practical work. It describes an operational development path:

Technical problem
→ patent and literature research
→ engineering knowledge graph
→ system reconstruction
→ generation of alternatives
→ simulation and validation
→ invention and patent strategy
→ data products and decision support

RCS can initially build this infrastructure as its own lead user.

Research required for developing and protecting new technical solutions is not treated as a one-time cost. It becomes part of a reusable engineering knowledge asset.

Each project can contribute:

additional system models,
improved ontologies,
validated relationships,
patent-family structures,
simulation results,
rejected alternatives,
synthetic engineering data.

This produces a reinforcing cycle:

$$
\text{better engineering knowledge}
\rightarrow
\text{better system generation}
\rightarrow
\text{better inventions}
\rightarrow
\text{better data}
\rightarrow
\text{better engineering knowledge}.
$$

The same infrastructure can then support:

internal R&D,
technology landscapes,
patent and competition analysis,
investment due diligence,
system-design projects,
licensing strategies,
specialised data products,
Graph-RAG applications.

Conclusion

Artificial intelligence does not need to converge on a single human-like form of general intelligence.

It can develop specialised forms of intelligence adapted to fundamentally different objects and environments.

Synthetic Systems AI is one such direction.

It treats systems—not documents—as the primary objects of intelligence. It combines structured engineering knowledge, generative search and simulation to understand existing architectures, create alternatives and evaluate their consequences.

Its purpose is not simply to describe possible futures.

It is to make possible systems:

explicit,
executable,
comparable,
testable,
protectable,
economically interpretable.

Synthetic Systems AI is intelligence for constructing the space of possible systems.

At Renner Complex Systems, we see this as a foundation for a new generation of engineering, invention and technology-intelligence infrastructure.