Advances and challenges for fluorescence nanothermometry

Fluorescent nanothermometers can probe changes in local temperature in living cells and in vivo and reveal fundamental insights into biological properties. This field has attracted global efforts in developing both temperature-responsive materials and detection procedures to achieve sub-degree temperature resolution in biosystems. Recent generations of nanothermometers show superior performance to earlier ones and also offer multifunctionality, enabling state-of-the-art functional imaging with improved spatial, temporal and temperature resolutions for monitoring the metabolism of intracellular organelles and internal organs. Although progress in this field has been rapid, it has not been without controversy, as recent studies have shown possible biased sensing during fluorescence-based detection. Here, we introduce the design principles and advances in fluorescence nanothermometry, highlight application achievements, discuss scenarios that may lead to biased sensing, analyze the challenges ahead in terms of both fundamental issues and practical implementations, and point to new directions for improving this interdisciplinary field. This Review on nanothermometry introduces the many types of nanothermometers and their cellular and in vivo applications, as well as best practices for accurate measurements.

A t the foundation of current nanothermometry is the fact that materials possess fluorescence properties that are strongly dependent on temperature, which can be used for remote and high-sensitivity thermal readouts at the nanoscale 1 . Such materials are referred to as (fluorescent) nanothermometers. Contactless nanothermometry with sub-microscale resolution has found numerous applications in different fields, including nanofluidics 2 , catalytic reactions 3 , microelectronics 4 and, perhaps most importantly, cell biology, preclinical research and diagnostics [5][6][7][8][9][10][11] . Nanothermometers have enabled challenging experiments such as the measurement of intracellular temperature or the early detection of tumors 12,13 . This potential has attracted numerous researchers to work on the design and synthesis of temperature-sensitive nanomaterials [9][10][11][14][15][16] . Advanced physicochemical synthesis routes have led to the discovery of new temperature-dependent properties in a range of different materials. A collection of nanothermometers with both high thermal sensitivity and preselected working wavelengths have been developed that have shown promise for biological research and in vivo diagnostics in proof-of-concept demonstrations.
Because the performance of nanothermometers employing different sensing strategies has been evaluated using a diverse range of assessment protocols and under different conditions, the potential use of fluorescence nanothermometry for either cell research or animal studies has not been fully demonstrated. Moreover, the complex in vivo environment can lead to biased or inaccurate sensing. In this Review, we survey mechanisms based on the photophysical properties and temperature-sensing strategies of fluorescent materials, evaluate the recent progress in using nanothermometers for intracellular temperature detection and in vivo diagnostics, and point out aspects of these approaches that can cause biased readouts. We discuss the challenges and opportunities in developing next-generation nanothermometer materials and state-of-the-art imaging techniques with the goals of improved spatial, temporal and temperature resolutions.

Design principles and advances of fluorescent nanothermometers
In the past decade, temperature-responsive fluorescence materials with biocompatible surfaces have been developed for noncontact intracellular and in vivo temperature monitoring. Organic species include fluorescent proteins (FPs), small organic compounds (dyes), lanthanide complexes and dye-doped polymeric nanoparticles. Inorganic nanoparticles include quantum dots (QDs), lanthanide-ion-doped nanoparticles (Ln-NPs), vacancy-containing nanodiamonds (NDs) and carbon dots (CDs). Their temperature-sensing features, for which the commonly involved indicators are defined in Box 1, are discussed in the next three sections. cells. Meanwhile, the inhomogeneous concentration of nanothermometers across the cell may lead to incorrect temperature values. The lifetime of these polymeric nanoparticles (② in Fig. 1a), which is prolonged from 4 ns to ~9 ns when the temperature is increased from 20 °C to 40 °C, can provide concentrationindependent readouts.
Fluorescence lifetime imaging microscopy has been widely used for nanothermometers. For example, silicon nanoparticles 18 show a decrease in lifetime from 23 ns to 8 ns as the temperature is increased from 0.5 °C to 60 °C, due to thermally favored non-radiative processes involving both vibrations and rotations of surface ligands. Temperature activation of non-radiative relaxation has also been observed in CDs 19 , in which the lifetime decreased from 11 ns to 5.3 ns as the temperature increased from 2 °C to 80 °C.
Temperature-induced spectral shifts (③ in Fig. 1a) are common in semiconducting QDs, in which the bandgap of the QDs changes with temperature according to Varshni's expression, E g (T) = E g0 − αT 2 /(T + β) (Fig. 1c) 20 . Unlike the intensity, the spectral shift is independent of the local QD concentration, allowing intracellular temperature sensing with a sensitivity of 0.16 nm K -1 using two-photon imaging of CdSe QDs 21 . Particle-size-dependent spectral shift has been further observed in CdTe QDs, with a sensitivity range of 0.2-0.8 nm K -1 (ref. 22 ).
Ratiometric design, by monitoring the intensity ratio between two emission or excitation peaks (④ in Fig. 1a), is prevalent due to its relatively high resistance to environmental interference. In Ln-NPs, by measuring the emission intensity ratio (I 2 /I 1 )), emitting from the thermally coupled energy levels E 1 and E 2 , the absolute temperature can be deduced according to the Boltzmann distribution equation I 2 =I 1 ¼ Ce �ΔE=k B T I (Fig. 1d). A full knowledge of the physical mechanisms underlying the thermal sensitivity makes it possible to check for possible artifacts and enables absolute temperature readouts, leading to what are known as primary thermometers. Typical thermally coupled energy-level pairs include Er 3+ ( 2 H 11/2 and 4 S 3/2 ) 23 , Nd 3+ ( 4 F 5/2 and 4 F 3/2 ) 24 and Eu 3+ ( 5 D 1 and 5 D 0 ) 25 . Sensing based on excitation intensity ratio has been demonstrated using green fluorescent proteins (GFPs), in which the ratio could be tuned by conformational changes of GFP-tagged proteins (Fig. 1e) 6 .
Fluorescence polarization anisotropy (FPA) (⑤ in Fig. 1a) is defined as a ratio related to the emission intensity collection from both the parallel and perpendicular polarizations with respect to the incident polarization. The measured FPA is closely related to molecular rotation arising from Brownian dynamics according to Perrin's equation 1/r = (1/r 0 )(1 + τ F /τ R ) (Fig. 1f). Small-molecule nanothermometers, such as GFPs and organic dyes, have this sort of dynamic FPA 26,27 , with increasing temperature resulting in smaller FPA due to a faster rotation of the molecules so that τ R = Vη(T)/k B T (Fig. 1f).
ESR and ODMR sensing strategies are used in vacancy-containing diamond nanothermometers (⑥ in Fig. 1a) 5,28 . Each nitrogen vacancy (NV) center has a ground-state spin triplet (m s = 0, ±1) (Fig. 1g). At zero magnetic field, the transition frequency (Δ) between the m s ¼ 0 where m s refers to spin projection) has a temperature dependence of dΔ dT ¼ �77 kHz K �1 I (ref. 5 ). This means that a rise in temperature shifts Δ to a lower frequency that can be optically measured with coherent control using microwave pulses. NDs can also be applied for all-optical temperature sensing by using the spectral shift of the zero-phonon lines of NV centers 29,30 or other defects (silicon-, germanium-and tin-vacancy centers) in the lattice [31][32][33] .
Progress in intracellular temperature sensing. Nanothermometers can be delivered into cells by endocytosis, chemical diffusion or microinjection for intracellular temperature measurements. They can also be targeted to label organelles, and the local temperature changes can be related to organelle functions. With the exception of those based on small molecules, most nanothermometers need to undergo surface modification to overcome their limitations in size, biocompatibility and surface hydrophilicity. Their functions are then validated by changing the temperature of the external medium, local heating by a laser or drug stimulation (Table 1). New knowledge of cellular biophysics, such as organelle-dependent thermogenesis and chemical-induced thermogenesis, has been obtained using fluorescent nanothermometers with high sensitivity to temperature variations, as summarized in Box 2.
Fluorescent protein nanothermometers are the most familiar to biologists and are advantageous in their superior capability for targeting to organelles and their ability to be genetically encoded. Expression-based FPs have shown their potential for temperature monitoring in the Golgi of HEK 293 cells 34 , the mitochondria of brown adipocytes (BAs) and endoplasmic reticulum (ER) of myotubes ( Fig. 2a) 6 , the mitochondria of HeLa cells 6,35,36 and temperature mapping of whole cancer cells 26,35 . Nevertheless, the reported thermal sensitivities of FPs remain relatively low (<2.82% K -1 ).

Box 1 | Glossary of terms used for fluorescent nanothermometers
In all fluorescence-based strategies, temperature is calculated from the measurement of diverse indicators, such as intensity, intensity ratio, peak position, polarization and lifetime. The value of these fluorescence indicators can be quantified as Q.
The rate of the change of Q associated with temperature T is defined as the absolute sensitivity: To compare the fluorescent nanothermometers with different indicators, the relative sensitivity is used: S r is comparable between different systems with the consistent unit of K −1 or % K -1 . The uncertainty comes from random variations in replicated independent measurements (type A) and systematic effects in the measurement process (type B).
Temperature resolution (T min ) is the smallest change in a temperature that causes a perceptible change in the fluorescence indicator. It is expressed by where σ is the s.d. of the parameter used for temperature determination. When a temperature reading is obtained by spectral analysis, σ is related to the signal-to-noise ratio in the emission spectrum. The σ, in turn, decreases with the integration time, so the temperature resolution depends on measuring time. The highest spatial resolution provided by a given fluorescent nanothermometer is determined by its size, which in most cases is <100 nm. In these conditions, the final spatial resolution of thermal images will be limited by the spatial resolution of the imaging system. When dealing with intracellular measurements, the use of conventional microscopes with high numerical aperture typically achieves diffraction-limited resolution.
The temporal resolution is determined by the integration (measuring) time. Fluorescent nanothermometers with low brightness require large integration times and hence provide poor temporal resolution.
Organic fluorescent dyes show both good temperature sensitivity and intracellular targeting ability. Localized temperature monitoring has been successfully used with systems including membrane and mitochondria of HEK 293 cells 17,34 , ER and mitochondria of BAs 37,38 and HCT116 cell spheroids 39 . However, the photostability of dyes varies depending on their structure. For example, the ER Thermo Yellow nanothermometers photobleach during imaging at a rate 40 of 1.5% min -1 , which requires data correction to sense the temperature accurately. ERthermAC can provide improved photostability 37 .
Use of a fluorescent nanogel with a diameter of ~50 nm for intracellular thermometry was demonstrated in 2009 41 . Through microinjection into the cytoplasm, this was used to monitor intracellular temperature with a resolution of 0.29-0.5 °C. However, the large size and low hydrophilicity of the molecules hindered their homogeneous dispersion throughout the cell and limited their use for intracellular temperature mapping. In 2012, Okabe and co-workers developed a linear polymeric nanoparticle with a reduced size of 8.9 nm and sufficiently hydrophilic residues 8 that it could be used to map intracellular temperature differences between nucleus, centrosome and cytoplasm, as well as local heat production from mitochondria (Fig. 2b). The development of cationic polymeric nanoparticles has allowed their easy cell internalization through incubation, even in yeast cells, which have cell walls 42 . The further design of a ratiometric cationic polymer has allowed the widespread use of nanothermometers in biological studies (Box 2) 43,44 .
Nanothermometers based on lanthanide complexes, made by embedding the complexes in PMMA with a cationic polymer at the surface to form nanoparticles, have an average hydrodynamic diameter of ~200 nm. Described as "walking nanothermometers" 45 , they are enclosed in endosomes or lysosomes after endocytosis and can be transported along microtubules. Tracking the position of single nanothermometers yielded a high spatial localization accuracy of 5.3 nm. The ratiometric design of nanothermometers revealed the heterogeneous heat production in a single cell after the application of ionomycin 46 .
A range of inorganic nanothermometers, including QDs, Ln-NPs, NDs, CDs and nanohybrids, have been reported for intracellular thermometry. Endocytosed individual QDs sense the temperature during Ca 2+ shock with a spatial localization accuracy beyond the optical diffraction limit 47 , though QDs appear only randomly in the cytoplasm (Fig. 2c). At the single-particle level, photobleaching of  Fig. 1 | typical temperature-sensing strategies and working mechanisms of different nanothermometers. a, Temperature-sensing strategies relying on ① emission intensity, ② lifetime, ③ peak position, ④ emission/excitation intensity ratio, ⑤ fluorescence polarization anisotropy and ⑥ electron spin resonance (ESR) or optically detected magnetic resonance (ODMR). b, The NNPAM unit in polymeric nanoparticles shrinks to release water molecules at elevated temperatures, increasing the emission intensity and lifetime. c, The temperature dependence of the bandgap of a QD (E g (T)) is described by Varshni's expression, where E g0 is the energy gap at 0 K, α is the temperature coefficient and β is close to the Debye temperature of the material. d, The electrons in the lower excited state (E 1 ) are thermally populated into the higher excited state (E 2 ) in Ln-NPs, so that the temperature dependence of the emission intensity ratio follows the Boltzmann distribution, where C is a constant, ΔE is the energy gap between the two excited states, k B is the Boltzmann constant and T is the absolute temperature. e, Temperature-dependent conformational change of GFP-tagged proteins, leading to a change in excitation intensity ratio. f, Molecular rotation arising from Brownian dynamics and described by Perrin's equation, where r is the polarization anisotropy, I || and I ⊥ are the intensities of the polarized fluorescence parallel and perpendicular to the incident polarization, r 0 the fundamental anisotropy, e the fluorescence lifetime, τ R the rotational correlation time, η(T) the dynamic viscosity of the medium, V the hydrodynamic molecular volume and k B the Boltzmann constant. g, At zero magnetic field, the ground-state spin m s ¼ ± 1 j i I sublevels of nitrogen-vacancy centers in nanodiamonds are split from the m s ¼ 0 j i I state by a temperature-dependent splitting Δ(T) due to thermally induced lattice strains. The spin states can be coherently manipulated using microwave pulses and efficiently initialized and detected using laser illumination. Panels b, e and g adapted with permission from ref. 8 , ref. 6 and ref. 5 , respectively; Springer Nature. up to 80% of QD655 has been observed within 5 minutes 48 . Yb 3+and Er 3+ -co-doped upconversion nanoparticles (UCNPs) with high photostability and detection contrast from negligible background have been reported to monitor the temperature of individual HeLa cells 23 . The challenges in using UCNPs include the specific labeling of organelles 49 and local heating due to the overlap between their excitation band and an absorption band of water. NDs allow mapping of the intracellular temperature gradient, and challenges with this approach include the requirement for invasive intracellular delivery-for example, silicon nanowire or needle delivery 5 -or long-term incubation (up to 12 h) 50 .

Reliability of intracellular nanothermometry.
Notwithstanding the rapid progress in intracellular nanothermometry, a critique of its reliability arises from the results of theoretical modeling 51 . Both the detected temperature heterogeneities (for example, Fig. 2b) and the chemical-stimulation-induced temperature rises (for example, Fig. 2c) in single living cells are too large compared to those (ΔT~10 −5 K) estimated from the conventional thermodynamic equation: where P, κ and L are the delivered heat power (in watts), the thermal conductivity (in watts per meter per kelvin) and the typical size of the heat source (in meters), respectively. The reasons for this discrepancy have been discussed in the contexts of (1) the validity of Eq. (4) in complex cellular systems 52,53 and (2) re-evaluation of cellular P, κ and L values that can be assigned to Eq. (4) 13,54-56 , but have not been fully clarified 57 . We expect that rational models will be required to enable a fuller examination the performance of nanothermometers against theoretical predictions.
Notably, further advances in this discussion will require full confidence in the intracellular temperature data. The reliability of intracellular fluorescent nanothermometers is becoming gradually established through the observation of reproducible significant temperature variations with various kinds of nanothermometers using different sensing strategies (Table 1) and even with recently developed nonfluorescent cellular thermometers such as thermocouples 58,59 , micro Si resonators 60 , thermistors 61 and Raman microscopes 62,63 . Despite this improvement in reliability, intracellular thermal images still require the use of a second method to ensure their validity. This could be achieved by simultaneously recording the physiological responses of other indicators, such as mitochondrial depolarization, oxygen consumption or extracellular acidification rate. The use of these physiological indicators to check the correctness of the thermal fluctuation values measured by fluorescent nanothermometers will require some maturation of current theoretical research 64 . Moreover, we anticipate that the use of a single nanothermometer with multiple readouts to indicate the same temperature value should lead to reliable intracellular temperature measurements. The effectiveness of this concept has been recently demonstrated in vivo 65 , but not yet for intracellular measurements.

Progress with in vivo temperature sensing in animal models.
In vivo applications of nanothermometers are still in their infancy.
Proof-of-concept demonstrations of nanothermometers in animal models include temperature-controlled photothermal therapy and diagnosis of tumors, inflammatory events and cardiovascular diseases. Inorganic nanoparticles showing near-infrared (NIR) excitation/emission features allow high penetration depth through biological tissues (Table 2). NIR light can penetrate hundreds of microns through bone tissue, enabling transcranial thermometry.
For semi-transparent organs, taking temperature readings is straightforward, as the same visible-emitting nanothermometers used for cellular thermometry are adequate in this context. Genetically encoded thermosensitive FPs have made it possible to study the response of Caenorhabditis elegans (Fig. 2d) and bacteria to heating stimuli 66,67 . A lanthanide complex thermometer and reference dye embedded in a polymeric matrix allowed two-dimensional (2D) ratiometric mapping of the body temperature of fly larvae (Fig. 2e) 68 and of the beetle Dicronorhina derbyana 69 . Intensity-based nanothermometers are of limited usefulness, however, in cases where the local concentration of nanosensors may fluctuate with time and be erroneously attributed to temperature variations.
Fluorescence nanothermometry in animal models has been primarily explored for real-time in situ temperature sensing during photothermal therapy (PTT) based on either multifunctional nanothermometers or hybrid nanostructures. Multifunctional nanothermometers from a single material display temperature-sensitive emission when excited by a single light source also used for therapy. Materials include Nd 3+ -doped LaF 3 nanocrystals 70 and PbS-based QDs. Nd 3+ -doped nanocrystals show a ratiometric NIR-I temperature-sensitive emission due to the thermal coupling between Stark sublevels. Their photothermal conversion efficiency can be tuned by adjusting the concentration of Nd 3+ ions. Higher Nd 3+ concentrations increase both the light absorption and the non-radiative photothermal conversion, but at the expense of fluorescence intensity, leading to a low signal-to-noise ratio in the temperature readout. The intensity-based PbS-based QDs nanothermometers emit in NIR-II with a better thermal sensitivity (see Table 2), but only in measuring temperature variations. The higher absorbance of QDs leads to lower laser power density requirements (1.5-2 W·cm −2 , compared with 4 W·cm −2 required for LaF 3 :Nd nanoparticles), although these remain well above the desirable power densities used with photothermal agents, such as gold nanorods or graphene-based nanomaterials 71 .
Two subunits for photothermal heating and independent temperature sensing can be combined to build hybrid nanothermometers. Although requiring a rather complex experimental setup, sometimes including two laser sources, this strategy allows the decoupling of heating and thermal sensing, as high photothermal conversion efficiencies can be achieved using nonfluorescent materials. One example is the nanocomposite formed by a NaLuF 4 :Yb, Er UCNP core for temperature sensing excited at 980 nm and a

Box 2 | Fluorescent nanothermometers contributed to biological studies
A small organic-molecule-based nanothermometer, MitoThermo Yellow, measured a mean mitochondrial temperature rise in HEK293 cells to be ~10 °C under full activation of respiration 17 . The DyLight549 molecular thermometer monitored the membrane temperature change during remote control of ion channels and neurons 34 . Brown adipocytes are known to regulate energy expenditure though non-shivering thermogenesis in humans and therefore are emergent targets for treating obesity. Different groups have focused on different molecules to activate brown adipocytes for thermogenesis. Fluorescent polymeric nanothermometers accurately monitored the effects of apoptosis signal-regulating kinase 1 (ASK1) 7 , the β3-adrenergic receptor agonist CL316.243 43 , the natriuretic peptide carperitide 44 and the ER-resident sensor PKR-like ER kinase (PERK) 112 on the thermogenesis of brown adipocytes. In addition, fluorescent polymeric nanothermometers have been applied to intracellular thermometry at the tissue level 113 , in which brain ischemia was associated with serious edema through remarkable temperature increases in brain tissue. Related examples can be found in brown adipocytes with a fluorescent polymeric nanothermometer 114 and in embryos of C. elegans with NDs 115 , respectively. photothermal carbon shell for PTT at a low laser power density of 0.3 W·cm −2 at 730 nm. Irreversible nanothermometers constitute another interesting option for thermometry during thermal ablation treatments 72 , where the post-treatment temperature sensing is unnecessary. Although these probes provide high relative sensitivities, their application in animal models remains unexplored.
Besides monitoring the progress of hyperthermia-based therapies, the use of in vivo nanothermometry has been explored in disease diagnosis. An increase in the local temperature is directly associated with inflammatory responses. A mouse model of inflammation has been used to evaluate the in vivo performance of hybrid nanothermometers that combined temperature-sensitive upconversion units based on triplet-triplet annihilation (TTA) with temperature-insensitive Nd-doped nanoparticles in a bovine serum albumin matrix 9 . These hybrid nanothermometers were capable of providing a subcutaneous thermal image of inflamed tissues (Fig. 2f). The major limitation of these hybrid systems is the great attenuation of visible light by tissues, which results in very low signal-to-noise ratios and the requirement of two beams for excitation, complicating the experimental set-up.
Other diagnostic applications include studying thermal dynamics by monitoring the thermal relaxation of tissues after external heating stimuli. Different NIR-emitting nanothermometers have been used to characterize this process in vivo in healthy specimens 73,74 and mouse models of ischemia/inflammation 75 and melanoma 76 . Ximendes and co-workers used core/shell LaF 3 :Er,Yb@LaF 3 : Yb,Tm nanocrystals with a 5% K -1 sensitivity at 20 °C for real-time recording of subcutaneous temperature 74 . Brain thermometry has also recently been demonstrated using NIR-emitting Ag 2 S QDs in a murine model of coma 77 .

Biased sensing
Certain temperature-independent factors can influence the fluorescence of a nanothermometer. Understanding them is key to avoiding artifacts and biased readouts.
Bias can arise from the power dependence of fluorescence signals. For example, ratiometric temperature sensing often involves detecting the emission bands of two materials or two emission bands of a fluorescence probe, I 1 and I 2 . If I 1 and I 2 emitted from the excited states E 1 and E 2 follow a one-photon linear population process and a two-photon upconversion process, respectively, emission intensity I 1 ∝ excitation power density, P, and I 2 ∝ P 2 , and the fluorescence intensity ratio (FIR) between I 2 and I 1 thereby becomes power dependent; that is, I 2 /I 1 ∝ P (Fig. 3a). The temperature readout can be biased due to an incorrect translation of the FIR, as the result of the different excitation power densities used in the calibration and the actual testing experiment. A difference in power density of one order of magnitude will cause a 30 °C temperature error (Fig. 3b).
Bias can also arise from the changes in refractive index values between different organelles (Fig. 3c). For instance, mitochondria have typical refractive indices between 1.4 and 1.42, and the nucleus has values of 1.355-1.365 78 . Nanothermometers located in a heterogeneous environment may behave differently in regard to electron transition probability and thus affect the emitted signals. For example, the difference in the refractive indices of DMSO and methanol, 1.51 and 1.36, can cause a type of Ln-NPs (refractive index 1.8) to display lifetimes of 26 ns and 35 ns when dispersed in DMSO and methanol, respectively 79 (Fig. 3d). Bias will occur in the temperature measurements of mitochondria and nucleus when such nanothermometers are used in conjugation with lifetime-based methods.
The complexity of biological tissues and their interactions with light creates several possible confounding variables, adding new levels of challenges for temperature sensing in living organs. Bias can be produced by fluorescence distortions caused by tissue autofluorescence (Fig. 3e). For example, the NIR emission from Ag 2 S nanothermometers partially overlaps with tissue autofluorescence under 800 nm excitation (Fig. 3f), and the overall spectrum changes when the ϕ value (the relative weight of autofluorescence to Ag 2 S emission) changes. If the temperature readout is obtained by analyzing the FIR of I 2 /I 1 or the peak position, the results will contain the autofluorescence contribution, which cannot be included in the calibration due to the variation between specimens. This otherwise can lead to readout errors of up to several kelvin.
Similarly, bias may also be caused by fluorescence distortions resulting from absorption and scattering effects due to the tissue composition (Fig. 3g), in which light is attenuated in a wavelengthdependent manner. The shape of the emission band of Ag 2 S nanothermometers obtained in vivo from liver differs by a large amount from that obtained in subcutaneous tissue 80 (Fig. 3h). If these distortions are erroneously attributed to temperature, that may lead to a temperature readout bias of up to 40 °C. Moreover, temperature affects the optical properties of tissues. At temperatures >45 °C, often used in PTT of tumors, heat will substantially affect the wavelength-dependent scattering coefficient and subsequently alter the emission ratios and introduce error in the measured temperature.
To alleviate these unwanted biases, we suggest carefully considering the different conditions used in establishing the calibration curves and the real testing scenarios. Some fluorescent materials have large absorption cross-section and low quantum yield, which converts a large fraction of absorbed energy to heat through Circled numbers indicates the type of temperature-sensing strategy, as categorized in Fig. 1a.
the non-radiative transition upon excitation 81 ; therefore, such self-heating should be evaluated during the calibration experiment. Nanothermometers with NIR excitation and emission within the optical transparency windows and a lifetime-based sensing modality are highly recommended for in vivo models. In these cases, special care should be taken to choose nanothermometers operating in spectral regions where the optical properties of tissues are wavelength independent, or at least known.

Challenges and emerging opportunities
In this section, we project that the emerging field of nanothermometry will involve material sciences and spectroscopy as means to develop a toolbox of nanothermometers with high sensitivity and reliability and to integrate new functionalities; interfaces with bioand nanochemistry to enable these nanothermometers to target the specific subcellular compartments; and adoption of new imaging technologies to super-resolve their distributions and dynamics. We further survey the list of ideal temperature resolution features needed for monitoring the metabolism of internal organs and suggest several recent advances that can potentially overcome the aforementioned challenges.
High brightness and relative sensitivity. Both high relative sensitivity, S r , and high brightness are required for the practical application of nanothermometers, as they result in high imaging contrast (signal-to-noise ratio) and high temperature resolutions (see Eq. (3) in Box 1), respectively. An S r of ~3% K -1 has been com-monly used to detect temperature variations of several kelvins (Tables 1 and 2). To improve the resolution to the sub-kelvin range, a higher S r is needed. An intrinsic difficulty with nearly all fluorescent materials to date is that high temperature often 'kills' their brightness. Fortunately, several recent reports have demonstrated the possibility of overcoming this limitation by using materials with positive temperature responses. We recently reported 14 that the multiphoton upconverted emissions from a type of sub-10-nm UCNP can be enhanced by up to 2,000-fold when the temperature is increased from room temperature to 453 K. A heterogeneous sandwich nanostructure combining thermally enhanced and quenched units has also been produced, resulting in a new type of ratiometric thermometer with a maximum S r of 9.6% K -1 at room temperature (Fig. 4a). Using a similar design principle, remarkable progress has been achieved with a TTA upconversion system (Fig. 4b) 9 . The simultaneous acquisitions of the orthogonal emissions at visible (thermally enhanced) and NIR (temperature independent) wavelengths from a nanocomposite of BDM&PtTPBP (organic unit) and β-NaYF 4 : 5%Nd (inorganic nanocrystal) through two independent laser excitations allow ratiometric sensing with a maximum S r of ~7.1% K -1 at ~22 °C. Further shifting both excitation and dual emission bands into the NIR region will be essential for deep-tissue applications.  nanoscale. Design and synthesis of hybrid materials to incorporate functional units could provide a real-time in situ feedback system to improve therapeutic accuracy, and the reverse control of temperature (refrigeration) will play a new role in investigating the effects of temperature on many physiological processes. Although efficient heating sources are relatively easy to find, including plasmonic, magnetic and carbon materials 10,82-84 and Nd 3+ -based lanthanide materials 85,86 as described above, cooling materials are rare. To date, only Yb 3+ doping has been identified as an effective refrigeration engine 87,88 . Both types of active nanothermometer-hybrid nanoheaterthermometer and nanocooler-thermometer-can be made using core-shell nanostructures. A plasmonic magnetochromic nanoheater has been developed for simultaneous local heating and thermometry measurement 89 (Fig. 4c). The heterogeneous nanodomes were made of 100-nm-diameter polystyrene beads partially coated with an alternating Au/Co multilayer. The plasmonic Au was responsible for heating, and the ferromagnetic Co responded to the applied magnetic field from a coil and drove the nanodomes to rotate. Local temperature sensing could be carried out by measuring the viscosity variation of the fluid; namely, water. This active nanothermometer showed a temperature resolution of 0.05 °C at room temperature. In a laser cooler made from Yb 3+ -doped LiYF 4 nanocrystal 87,88 , using an appropriate laser-trapping wavelength of 1,020 nm, the anti-Stokes transition of Yb 3+ can absorb heat from the local environment to lower the local temperature (Fig. 4d). By co-doping Er 3+ ions, ratiometric fluorescence thermometry allowed the qualitative reading of the cooling effect when the trapping irradiance increased. Remarkably, Roder and co-workers demonstrated that 2%Er 3+ and 10%Yb 3+ co-doped LiYF 4 can undergo either laser refrigeration (ΔT = −4.9 ± 2.8 °C) at λ = 1,020 nm or heating (ΔT = 21.8 ± 10.11 °C) at λ = 975 nm 87 .

New sensing modalities.
A multi-modality thermometer allows cross-validation of the temperature readout. This has been recently demonstrated by performing three independent measurements on a single trapped, rotating upconversion microparticle (Fig. 5a) 90 . The internal temperature of the microparticle was raised by the trapping laser owing to the absorption of the Nd 3+ dopants, which can be measured according to both the internal and external degrees of freedom, through upconverted fluorescence, rotation rate and Brownian dynamics of the particle. Both rotation rate and trap stiffness methods achieved the same temperature sensitivity (2.0% K -1 ), which was three times better than that measured by fluorescence (0.66% K -1 ). The thermal loadings measured from the fluorescence and rotation rates were in close agreement, whereas the stiffness method yielded lower values, with the difference increasing at higher trapping power (Fig. 5b). The development of ratiometric nanothermometers using temperature-responsive lifetimes can minimize the complex tissue interference seen with in vivo thermometry, as a single emission wavelength used. A hybrid nanoparticle made of UCNPs, PbS QDs and a surface layer of SiO 2 (Fig. 5c) 91 allows lifetime thermometry, taking advantage of the large difference in the ~800-nm band emission lifetime of Tm 3+ -doped UCNPs at microsecond and PbS QDs at nanosecond time scales (Fig. 5d). This strategy did not lead to an obvious deviation in the intensity ratios of the nanothermometers at different tissue depths (0-3 mm) (Fig. 5e) while achieving a relative sensitivity of ~5.6% K -1 and a temperature resolution of ~0.5 °C at ~45 °C. The ratio tissue blocking (1-3 mm) /ratio true (0 mm) was close to 1 at different temperatures for the new time-resolved thermometry, indicating its immunity to tissue interference (Fig. 5f). By contrast, conventional ratiometric thermometry suffered from obvious depth-dependent interference.
Real-time high spatiotemporal resolution. The uncertainty in intracellular temperature sensing is usually due to the complexity of the surrounding environment and the dynamic structural changes of organelles, which require high spatiotemporal resolution, accessible with state-of-the-art real-time super-resolution imaging techniques.
The mitochondrion is a commonly studied organelle in both temperature-sensing and super-resolution microscopy because of its dynamics in local temperature generation and abundant morphological structure changes. In 2018, Chrétien and co-workers found that mitochondrial temperatures were some 10 K above that of the surrounding water bath by using the MitoThermo Yellow thermometer (Fig. 6a) 17 . A major challenge in experimentally validating these results is to localize nanothermometers in relation to mitochondria in real time 92 . This requires both specific targeting and immobilization of nanothermometers to specific sites within mitochondria and super-resolution imaging technologies to resolve them.
Typically, dye-based nanothermometers are used to target mitochondria 93,94 through electrostatic interaction. However, this interaction is too weak to anchor the probes when the membrane potential changes, for example, upon exposure to CCCP, which causes the nanothermometers to move away from the mitochondria. Applying the lesson of immobilized probes 95 , such as the MitoTracker series with their benzyl chloride group, covalent binding of chemically reactive groups could be implemented to immobilize nanothermometers. Huang and co-workers have developed a fixable nanothermometers, Mito-TEM 96 , using the benzaldehyde group to conjugate with the amino group of the protein and rhodamine B as the thermal sensing unit (Fig. 6b).
The resolution of conventional diffraction-limited optical microscopy is insufficient to image submitochondrial structures, as the crista-to-crista distance is often <100 nm 97 . The latest development of structured illumination microscopy based on Hessian matrices (Hessian-SIM) allowed a large-field-of-view observation of mitochondrial inner membrane dynamics with high spatiotemporal resolutions of 88 nm and 188 Hz (Fig. 6c) 98 . The recently developed highly photostable probes have achieved spatial resolutions <60 nm in imaging the dynamic structure of the mitochondrial cristae in living cells through the use of stimulated emission depletion (STED) nanoscopy [99][100][101] .
3D, reliable, real-time in vivo monitoring with resolution <0.1 °C. Temperature variations associated with in vivo physiological processes are ~1 °C (Fig. 7a), which requires the current thermal resolution of 0.5 °C to be lowered to 0.1 °C to accurately monitor epilepsy or acute cardiovascular accidents. The achieved temporal resolution of 2 s is also too slow to measure single-cell hyperthermia (Fig. 7b) or the fast temperature jumps that occur during epilepsy (Fig. 7c). To date, nanothermometers have been successful in providing real-time thermal feedback only during tumor hyperthermia treatments (Fig. 7d). In vivo thermal monitoring of other diseases or biological processes will require a reduction of the time resolution by at least one order of magnitude. Achieving better thermal resolution implies a requirement for reduced s.d., σ (see Eq. (3) in Box 1). As increasing the integration time during fluorescence measurements would harm the time resolution, the need for high-brightness nanothermometers is critical as a means to achieve high signal-to-noise-ratio fluorescence data (reduced σ) toward the resolution target of <0.1 °C and 0.1 s. The recently developed Ag 2 S@AgCl nanothermometer, fabricated through ultrafast photochemistry after the conventional wet-chemistry synthesis, is a promising candidate because of its significantly improved brightness 102 .
Improved thermal resolution and reliability can be achieved by combining multiparametric nanothermometers with neural networks (NN) or deep-learning techniques 103 . Spectral-shape-based NN (SP-NN, input neurons N in = 62) and multi-band NN (N in = 5) provide much higher reconstruction reliability, with temperature resolutions of 0.27 K and 0.25 K, as compared with integratedintensity-and peak-intensity-based NN (IP-NN, N in = 2) (Fig. 7e,f) 104 . The benefits of the neural networks have been demonstrated in fluorescence nanothermometry 104,105 but have yet to be applied to in vivo measurements.
Lifetime-based fluorescence thermometry can reduce the distortions caused by complex tissue, as fluorescence lifetime is not affected either by tissue scattering and absorption or by the concentration of thermometers. The time-gated imaging approach, based on a tunable delay between excitation and acquisition (Fig. 7g), works well for nanothermometers with lifetimes of >100 ns. Measuring shorter lifetimes, in the tens of nanoseconds, would require expensive picosecond-nanosecond laser pulses (Fig. 7h). Alternatively, phase-modulated excitation would allow lifetime imaging by analyzing the phase delay between excitation and emission intensity at each pixel (Fig. 7i), in analogy with the technique used in fluorescence lifetime imaging microscopy (FLIM). The latter approach may be suitable for preclinical lifetime imaging in vivo, which requires simple and low-cost excitation sources, fast cameras and algorithm development.
The desirable approach of 3D thermal imaging will further require the in-depth localization of single nanothermometers, with their temperature-responsive spectroscopy data, to be fully captured in real time. On the flip side, as the magnitude of spectral distortions depends on the tissue thickness (Fig. 7j), the analysis of spectra generated by nanothermometers opens a venue for assessing their depth. A similar approach has been demonstrated ,--(.) by using the thermal emissions of objects to extract their depthdependent temperatures 106 . Taking these approaches together, it should be possible to build a 3D functional imaging map, which also suggests the need for nanothermometers with a high lifetime thermal sensitivity and a broad emission band (so that spectral distortion is evidenced).
With the advent of reliable, high-resolution nanothermometry methods, future biological study methods will be transformed from structural into functional imaging of intra-and intercellular temperature dynamics. These improvements will enable basic biological investigation into the role of thermal fluctuations in living systems and allow real-time monitoring of the progression of diseases such as cancer and neurological disorders that are marked by significant alterations in metabolic activity and/or vascularization that could be observed as temperature changes.