Introduction
Science has often described scientific practice as a trichotomy of three fundamental elements:
- Theories: build a conceptual framework to explain and predict phenomena;
- Observations: collect data about the world;
- Experiments: allow controlled interventions on a physical system to test hypotheses, laws and models.
However, in many cases, we cannot directly manipulate the system of interest (e.g., gravitational waves, black holes, exoplanets and etc.), nor can we observe the phenomenon "directly" without complex chains of instrumental mediation.
Therefore, scientific practice incorporates tools that fall outside the theory-observation-experiment trichotomy: from thought experiments to computer simulations. These tools share a common trait: they allow us to exercise what we call surrogate reasoning, that is, a reasoning about a "substitute" (mental, mathematical, or computational), to reach plausible conclusions about something else (an object, a process, or a theory).
Sometimes, science goes beyond the traditional, because, the world we want to understand is often non-manipulable, not directly observable, or too complex to be observed without intermediate tools.
Why Scientific Practice Goes Beyond the Trichotomy
The traditional trichotomy works well when the ideal conditions for scientific method are met: we can intervene in the system, repeat the test, isolate variables, produce relatively transparent observations, and compare them with theoretical predictions. But there are at least three systematic reasons why these conditions are not always available.
Experimental Impossibility or Impracticability
Many systems are beyond the reach of human intervention: cosmic events, phenomena on enormous scales, or processes too dangerous to manipulate. Even when an experiment would be possible in principle, it may be economically unsustainable or require unavailable technologies. Here, science cannot do without indirect tools.
Heavily Mediated Observations
Many data are not simple readings of the world, but the output of extraction and transformation procedures (filters, reconstructions, calibrations, statistical inference). In these cases, speaking of observation as "pure raw data" is quite misleading.
Theoretical and Mathematical Complexity
A theory can be well-formulated yet not analytically calculable. Without computational tools, we would be unable to explore the consequences of the theory or produce usable predictions.
These three elements explain why scientific practice develops and uses additional tools. In particular, computer simulations exist in a hybrid zone: they resemble experiments because they allow exploration, manipulation (on parameters), and the production of "data-like" outputs; but they also resemble models, because they are surrogates for the real system and depend on theoretical assumptions. This hybridization is crucial to understanding their epistemic value and the philosophical problems they generate.
Gravitational Waves and Computer Simulations
Gravitational waves are ripples in space-time predicted by general relativity. The idea of detecting them might seem like a simple observational exercise: build a detector and measure. In reality, the process is much more complex and perfectly illustrates why the traditional trichotomy, alone, are not enough.
Interferometric detectors (such as LIGO and VIRGO) measure tiny variations in the length of the interferometer's arms. The expected signal is extremely weak and immersed in complex noise (seismic, thermal, instrumental, and etc.). Consequently, the question is not just "what did the instrument measure?" but "how do we distinguish a real physical signal from a fluctuation or an artifact?" and "how do we assign a precise physical interpretation to the signal?"
This is where simulations come into play in a crucial way. They serve to:
- Build theoretical templates (waveforms) for different physical scenarios (black hole mergers, neutron stars, and etc.);
- Perform detection through matched filtering: compare data with expected templates and search for statistically significant matches;
- Perform parameter estimation: once a signal is found, estimate physical parameters and uncertainties by comparing the data with families of waveforms, obtained via simulation;
- Assess robustness and significance: build background noise, estimate false positive rates and verify that the event is not caused by known bugs.
A key point is that simulation is not just "post-processing": it is an integral part of the identification criterion of a gravitational wave.
Main Philosophical Problems
Problem of the Theoretical Dependence of Evidence
The identification of the signal depends on templates produced by numerical modeling. It, therefore, seems that the evidence is profoundly theory-based. This raises the question: "are we really observing something in the world, or are we finding in the data what the theory tells us to look for?"
Problem of Justification and Risk of Circularity
If the signal is recognized by comparison with simulated waveforms, is there a risk of circularity: "does the theory produce the templates, and do the templates confirm the theory?"
Problem of the Reliability of Simulations
Numerical simulations require approximations and assumptions about initial conditions. How can we ensure that the outputs are reliable and not computational artifacts?
Problem of the Epistemic Status of Simulations
A simulation produces data-like outputs, but they are not data from the world: they are data from the model. In what sense, then, does it contribute to empirical knowledge? Is it a form of experiment? An extension of theory? A model of an experiment? This point is crucial because it determines how we evaluate the evidential strength of the discovery.
Two Approaches to Addressing the Philosophical Problems
Approach A: Simulations as a Form of Experiment
An immediate way to understand computer simulations is to compare them to experiments. Even if we don't directly intervene in the target system, we can still conduct experiments on its computational counterpart: we set initial conditions, evolve the system according to rules drawn from theories, observe the output, and then repeat the procedure by changing parameters and hypotheses. In this sense, the simulation is a virtual laboratory: intervention consists of varying inputs and assumptions; observation consists of interpreting results, often treated as data.
This perspective clearly explains the operational nature of simulations in contemporary science. Scientists use them not only to calculate theoretical consequences, but also to explore spaces of possibility: they perform repeated runs, compare outputs, look for patterns, study sensitivity to small variations, and sometimes encounter dynamics that are unexpected compared to initial expectations. Calling them "experiments in silico", therefore, captures an important aspect of the practice: the iterative, exploratory, and procedural aspects.
Approch B: Simulations as Surrogate Tools that Model Possible Experiments
A second approach starts from the idea that simulations resemble experiments in many ways, but do not coincide with them. What we call "intervention" in a simulation is not an intervention on the target, but rather on a representational system. Similarly, "observation" concerns outputs generated, according to a prior defined rules and assumptions. Consequently, simulation is best understood as a practice that stages possible experiments: it allows activities analogous to intervention and observation, but in a constructed environment, which provides direct information about what is true in the model and indirect information about the world.
The advantage of this perspective is that it preserves an important conceptual distinction: the simulation output is not automatically empirical data. It becomes, epistemically, relevant only when we can argue that the model captures relevant aspects of the target system, and that the errors introduced are understood and controlled. This also makes it clear how the scientific community treats simulations: as tools to be verified, compared, calibrated, and stressed with sensitivity analyses and cross-checks.
Why I Prefer Approach B
The case of gravitational waves makes it clear that discovery is not the result of immediate observation. Interferometric detectors measure tiny variations in high-noise conditions. Detection is an inferential chain in which raw data becomes evidence thanks to filters, statistical methods, and above all, expected waveforms obtained from modeling and numerical simulations. The central epistemic question is: "why we can say that a certain structure in the data is better explained as a gravitational wave rather than as noise or an artifact?
Approach B responds naturally: the simulation builds a set of physically possible scenarios (given a theoretical framework) and establishes what we should observe if the world realized one of these scenarios. The match between data and template is not an automatic proof, but a step that gains strength only within a set of controls. And this is exactly how the LIGO/Virgo works: reliability does not come from a single model or a single pipeline, but from a system of convergent checks.
For example, multiple and independent pipelines are used; the convergence of results is an indicator of robustness. Significances are estimated by comparing candidate events against an empirical background of noise, quantifying how unlikely it is to obtain a similar result by chance. Injection tests are conducted, inserting simulated signals into the data to verify that the pipelines recover them and to evaluate sensitivity and error rates. Finally, different families of waveform models are compared to understand how much parameter estimates depend on modeling choices. All these practices make sense if simulation is seen as a mediator between theory and data, not as an experimental substitute for the world.
Why I DON'T Prefer Approach A
The "simulations as experiments" approach risks to skip a key distinction: in real experiments, the system can behave in ways that exceed or contradict the model; In a simulation, what happens is bounded by the implemented rules and modeling assumptions.
In the case of gravitational waves, this is crucial, because the simulated output can match a structure in the data even for false reasons: instrumental glitches, rare statistical coincidences, biases introduced in template selection or filtering. Precisely for this reason, the scientific community insists on noise controls and instrument calibration. Interpreting simulations as experiments, risks promoting them to sources of direct empirical evidence, confusing consistency with the model, with confirmation of the phenomenon in the world.
Theoretical dependence does not imply circularity
The concern remains: "If I use general relativity to generate templates and then find a match in the data with those templates, aren't I just confirming what I assumed?" Approach B clarifies why this is not a circularity. The match is not guaranteed: the inferential chain is fallible and could fail at many points (noise, numerical errors, approximations, incomplete models). Credibility arises from the convergence of many conditions: consistency of the signal shape, consistency between detectors, significance with respect to the background, stability of results when pipelines and models vary. In this sense, theory, simulation, instrument, and data do not form a closed circle, but a mutual control loop: theoretical dependence is inevitable in highly mediated contexts, but can be epistemically virtuous when accompanied by robust community practices of validation and comparison.
Conclusion
Contemporary science shows that the traditional trichotomy, while useful, is insufficient to describe real-world practice. In many fields, knowledge arises from tools that act as epistemic mediators: thought experiments, models, and simulations. Computer simulations, in particular, have become indispensable when the target is too distant, too large, too complex, or too difficult to manipulate.
The case of gravitational waves highlights the most philosophically interesting aspect of this phenomenon: the discovery does not consist of a "direct observation", but of a network of inferences supported by simulations, statistical methods, control procedures, and collective validation practices. The approach that interprets simulations as models of possible experiments is more appropriate: it recognizes the epistemic power of simulations without confusing them with experiments on the world. The "simulations as experiments" approach, however, risks attributing excessive empirical authority to simulations, and obscuring their dependence on modeling assumptions and computational implementations.
Understanding science today, therefore, means understanding the traditional trichotomy, but also these intermediate tools and how, in community practice, they are controlled, justified, and integrated into the production of evidence.