Differential scanning calorimetry (DSC) heats a small sample and an empty reference pan on the same temperature program and records how much extra heat the sample absorbs or releases. When the sample melts, crystallizes, or relaxes, the trace shows a peak. That trace is not the sample's own heat release, though. It is that heat release after it has passed through the pan, the sensor, and the furnace, each of which delays and smooths it. Speech scientists face a structurally similar problem. A recorded voice is a glottal excitation, the "source," shaped by the vocal tract, the "filter." They separate the two by fitting both at once, with a physically constrained model for each, and resynthesizing the signal to check the fit.
We propose to apply that analysis-by-synthesis approach to DSC. The trick that makes the separation possible here is varying the heating rate. The instrument filter acts in time and does not depend on how fast the sample is heated. The sample's kinetics act in temperature and shift in a predictable, model-constrained way as the heating rate changes. Fitting one shared, temperature-dependent instrument filter to a series of runs at three to five heating rates should therefore separate instrument lag from intrinsic kinetics without special hardware or modulated temperature programs. The output would be sharper, rate-invariant heat-release curves and less biased transition temperatures.
Core Claim
A DSC thermogram can be modelled as an intrinsic heat-release source passed through a slowly temperature-varying thermal filter, with feedback from the source onto the sample's actual temperature. Jointly inverting a heating-rate series against this model should recover transition temperatures, peak splits, and kinetic parameters that agree across heating rates more closely than three alternatives: phenomenological peak fitting, single-time-constant lag correction calibrated on a reference metal, and extrapolation of peak temperatures to zero heating rate.
Two ingredients are new relative to anything in our search. The first is that the filter is estimated blindly and shared across rates rather than calibrated in a separate run. The second is the forward simulation of the coupling between latent heat release and sample temperature. A linear inverse filter cannot represent that coupling.
What the Two Directions Contribute
On the calorimetry side, the problem is well documented. Overlapping melting, cold crystallization, and relaxation events are routinely separated by fitting Gaussian- or Pearson-type peaks after the measurement. Li et al. used this to split dual melting peaks in neodymium butadiene rubber blends. They also note that critical temperatures from such fits disagree with each other across heating rates and sample preparations. That disagreement motivates our work, because it is the symptom we expect lag-induced smearing to cause. Zhang et al. fit the Avrami crystallization equation to isothermal traces of the same rubber family. Avrami fitting presupposes a clean, isolated peak, so it constrains our design: the source model must be kinetic, not a free-form shape.
Two existing approaches to the same problem rely on extra experimental effort rather than computation. Bechgaard et al. superimpose a sinusoidal temperature oscillation, a technique called temperature-modulated DSC, to split reversing heat capacity from non-reversing kinetics in silicate glasses above 700 °C. It lengthens runs, rests on linear-response assumptions, and showed a systematic offset in fragility (how steeply viscosity changes near the glass transition) against viscosity data. Sherman et al. add an optical microscope inside the furnace that detects crystal changes through second-harmonic generation. This resolves individual particle transformations, but only for crystals lacking a centre of symmetry and only in open pans. Neither approach separates the instrument's transfer function from the sample's kinetics, which is the gap we target.
On the speech side, we reuse methods rather than conclusions. Yu and Fazekas built a sample-wise, time-varying recursive filter that can be differentiated end to end inside an analysis-by-synthesis loop. They showed that frame-wise approximations smear fast dynamics. We adopt their recursive-filter gradient approach because our filter also changes as temperature sweeps.
Freixes et al. pair glottal inverse filtering with a low-dimensional parametric source model and a physical model of the vocal tract. They also flag source-tract acoustic interaction as a limitation of linear theory. We take their template of a constrained parametric source paired with a physically grounded filter, and we treat their interaction caveat as the analogue of latent-heat self-heating, which we model explicitly.
Choi et al. show that leakage between source and filter is the central failure mode of decomposition, and that strong separate priors reduce it. We borrow the evaluation idea of measuring leakage directly.
One tension from the speech side shapes the filter design. Speech filters are all-pole resonators, but heat conduction is diffusive and has no resonance. Carrying the speech filter over unchanged would produce non-physical ringing near the step change in heat capacity at a glass transition. Our filter is therefore restricted to real, negative poles: a cascade of first-order lags that cannot ring.
Research Program
The forward model for a run at heating rate $\beta$ is:
$$\Phi_\beta(t) = \big(h_{T}\ast[\dot Q_\beta + C_p(T_s)\,\dot T_s]\big)(t) + b(t), \qquad T_s(t) = T_0 + \beta t - R\,\Phi_\beta(t).$$
Each symbol is defined as follows. $\Phi_\beta$ is the measured differential heat flow. The filter $h_T$ is a cascade of $K\le 3$ first-order lags whose time constants vary smoothly with programmed temperature. $\dot Q_\beta = \Delta H\, d\alpha/dt$ is the intrinsic heat-release rate. $\alpha$ is the fraction of material converted, which evolves as $d\alpha/dt = k(T_s)\,f(\alpha)$ with an Arrhenius rate $k$ and a reaction model $f$, such as an Avrami or autocatalytic form. $C_p$ is the sample heat capacity. $T_s$ is the actual sample temperature. $R$ is the pan-to-sensor thermal resistance. $b(t)$ is a low-order baseline.
The second equation is the source-filter interaction. Latent heat drags the sample temperature away from the program, which in turn changes the kinetics. This is why we simulate forward rather than invert linearly.
Overlapping events are represented as a small mixture of kinetic sources with nonnegative enthalpies. The number of sources is chosen by held-out-rate prediction error rather than by eye.
Fitting minimizes squared resynthesis error summed over all heating rates. The filter parameters, $R$, and the baseline are shared across rates within a sample. Kinetic parameters are shared across both rates and samples of the same material. Gradients pass through the recursive filter and an ODE solver, following the sample-wise differentiation strategy of Yu and Fazekas.
We will also characterize identifiability analytically. In the temperature domain, filter smearing grows roughly linearly with $\beta$, because a fixed time constant spans $\beta\tau$ kelvin. Kinetic peak shifts instead grow roughly logarithmically, following the Kissinger relation between peak temperature and heating rate. These different scalings are what make the joint problem well-posed. We will show where they become indistinguishable, for example at a narrow range of heating rates or very fast kinetics.
Baselines
The naive reference is the raw thermogram analysed as recorded, with onset and peak temperatures and areas read directly. Our method must at least match its peak areas, since enthalpy is conserved by any lag. It must also improve its heating-rate consistency.
The strongest phenomenological competitor is multi-peak fitting of the kind used by Li et al., using skewed Pearson-type peaks on a fitted baseline. We run it on the identical traces with the same number of components our method selects. It gets the same optimizer budget and multiple random restarts.
The strongest physics-based competitor is classical single-time-constant lag correction. The time constant is calibrated from a sharp reference-metal melt, and the sample temperature is corrected by the resistance times the heat flow. This is standard calorimetry practice, but no paper in our search reports it, so we have no cited number. We give it the same calibration information our method would otherwise have to infer.
For the kinetic question, the competitor is the conventional Kissinger-style extrapolation of peak temperature to zero heating rate, combined with Avrami fitting as in Zhang et al..
Two ablations isolate our ingredients. In the first, we replace the blind, shared, temperature-varying filter with the calibrated fixed single lag, which tests whether blind multi-rate estimation matters. In the second, we drop the temperature-feedback term, which tests whether modelling the interaction matters. We also run an all-pole, resonance-allowed filter to measure the ringing the file warns about.
On reported numbers, no reported result was found on Lacuna for any DSC deconvolution-accuracy metric, nor for lag-corrected transition temperatures. The quantitative anchors available are material values, not method scores: a critical crystallization temperature of −32.5 °C for dual-peak onset and equilibrium melting points of about 0 °C (pure), −4.579 °C (50 PHR oil), and −6.579 °C (60 wt.% copolymer) from Li et al., and a crystallization activation energy of −264.3 kJ/mol for pure rubber from Zhang et al.. Beating a baseline therefore means three things:
- a statistically lower spread of recovered transition temperatures and peak-height ratios across heating rates on the same material;
- lower error against ground truth on simulated data;
- lower prediction error on a held-out heating rate.
These criteria apply at matched parameter counts.
Why not just…?
Why not just correct thermal lag with a calibrated time constant? This is the most serious threat. For a single sharp transition in a well-packed pan, it may do nearly as well. Our case rests on three situations where it should not. The effective time constant changes with sample packing and heat capacity, so a metal calibration transfers imperfectly. The filter drifts with temperature across a wide scan. Latent heat feedback distorts strongly exothermic crystallization. The first ablation settles the question directly. If the calibrated fixed lag matches our method on the rubber blends, we will report that the contribution reduces to removing calibration runs. We will not claim a sharpening gain in that case.
Why not just run temperature-modulated DSC? Bechgaard et al. show that it separates reversing from non-reversing heat. However, it multiplies run time, assumes linear steady-state response, and still needed empirical correction against viscosity-derived fragility. More fundamentally, it splits the signal by reversibility, not into sample versus instrument, so lag still smears the non-reversing part. Where modulated runs are available, we will compare their non-reversing component against our recovered source on the same material.
Why not just fit peaks or extrapolate to zero heating rate? Peak fitting is the source of the heating-rate inconsistencies Li et al. report, because its shapes absorb lag. Extrapolation to zero heating rate handles peak position but not peak shape or overlap. It also assumes linearity in the rate that our scaling analysis predicts fails. Both are run as baselines, so the comparison is empirical rather than argued.
Why not just add in-situ optics? The Sherman et al. approach works only for crystals lacking a centre of symmetry and only in open pans. Our method targets the sealed-pan, centrosymmetric, and amorphous cases it cannot reach. It is complementary rather than a replacement.
Proposed Experiments and Decisive Checks
The first stage is simulation with known ground truth. We use finite-element heat-conduction models of pan, sample, and sensor with varied contact resistance, sample mass, and temperature-dependent conductivity. These are driven by Avrami and autocatalytic sources, including two overlapping melting populations that mimic the rubber dual peaks. Metrics are source-curve error, transition-temperature error, kinetic-parameter error, and leakage. Leakage is the correlation between recovered filter time constants and true kinetic parameters across simulations, adapted from the speech notion of source-filter leakage in Choi et al..
The second stage is real calibrants: a pure metal melt run at five heating rates and three sample masses. A correct method should return a rate-independent, mass-independent onset and a near-delta source. This is the cleanest real-data check that blind filter estimation works.
The third stage replicates the neodymium butadiene rubber blend protocols of Li et al. and Zhang et al. at several heating rates. We will check whether our recovered dual-peak temperatures and height ratios become rate-invariant where theirs drift. We will also check whether the onset of the second peak near −32.5 °C sharpens, and how lag-corrected activation energies compare with their uncorrected values.
The decisive test is held-out-rate prediction. The model is fit on all but one heating rate and must predict the omitted thermogram. A model that has truly separated instrument from sample should extrapolate. A model that has merely absorbed lag into its peak shapes, as peak fitting does, should not.
A secondary test uses very thin, low-mass, slow-rate runs as a near-lag-free reference. Our recovered source from thick, fast runs should match it more closely than any baseline does.
References
- The Crystallization and Melting Behavior of Neodymium-Based Butadiene Rubber Blends
- Study on the Crystallization Behavior of Neodymium Rare-Earth Butadiene Rubber Blends and Its Effect on Dynamic Mechanical Properties
- Temperature-modulated differential scanning calorimetry analysis of high-temperature silicate glasses
- Structural analysis by stochastic differential scanning calorimetry
- Differentiable Time-Varying Linear Prediction in the Context of End-to-End Analysis-by-Synthesis
- Glottal inverse filtering and vocal tract tuning for the numerical simulation of vowel /a/ with different levels of vocal effort
- DDDM-VC: Decoupled Denoising Diffusion Models with Disentangled Representation and Prior Mixup for Verified Robust Voice Conversion